{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"},"kaggle":{"accelerator":"none","dataSources":[{"sourceId":49349,"databundleVersionId":5447706,"sourceType":"competition"},{"sourceId":71885,"databundleVersionId":8143495,"sourceType":"competition"},{"sourceId":5373920,"sourceType":"datasetVersion","datasetId":3117886},{"sourceId":3736,"sourceType":"modelInstanceVersion","modelInstanceId":2663},{"sourceId":3840,"sourceType":"modelInstanceVersion","modelInstanceId":2742},{"sourceId":3846,"sourceType":"modelInstanceVersion","modelInstanceId":2747},{"sourceId":3735,"sourceType":"modelInstanceVersion","isSourceIdPinned":true,"modelInstanceId":2662}],"dockerImageVersionId":30461,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"## Baseline submission\n\nA notebook to generate a valid submission. Implements three local feature/matcher methods: LoFTR, DISK, and KeyNetAffNetHardNet.\n\nRemember to enable a GPU accelerator and disable internet access, then press \"submit\" on the right pane.","metadata":{}},{"cell_type":"code","source":"# General utilities\nimport os\nfrom tqdm import tqdm\nfrom time import time\nfrom fastprogress import progress_bar\nimport gc\nimport numpy as np\nimport h5py\nfrom IPython.display import clear_output\nfrom collections import defaultdict\nfrom copy import deepcopy\n\n# CV/ML\nimport cv2\nimport torch\nimport torch.nn.functional as F\nimport kornia as K\nimport kornia.feature as KF\nfrom PIL import Image\nimport timm\nfrom timm.data import resolve_data_config\nfrom timm.data.transforms_factory import create_transform\n\n# 3D reconstruction\nimport pycolmap","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print('Kornia version', K.__version__)\nprint('Pycolmap version', pycolmap.__version__)\n\nLOCAL_FEATURE = 'KeyNetAffNetHardNet'\ndevice=torch.device('cuda')\n# Can be LoFTR, KeyNetAffNetHardNet, or DISK","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def arr_to_str(a):\n    return ';'.join([str(x) for x in a.reshape(-1)])\n\n\ndef load_torch_image(fname, device=torch.device('cpu')):\n    img = K.image_to_tensor(cv2.imread(fname), False).float() / 255.\n    img = K.color.bgr_to_rgb(img.to(device))\n    return img","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Image Pairs","metadata":{}},{"cell_type":"code","source":"# We will use ViT global descriptor to get matching shortlists.\ndef get_global_desc(fnames, model,\n                    device =  torch.device('cpu')):\n    model = model.eval()\n    model= model.to(device)\n    config = resolve_data_config({}, model=model)\n    transform = create_transform(**config)\n    global_descs_convnext=[]\n    for i, img_fname_full in tqdm(enumerate(fnames),total= len(fnames)):\n        key = os.path.splitext(os.path.basename(img_fname_full))[0]\n        img = Image.open(img_fname_full).convert('RGB')\n        timg = transform(img).unsqueeze(0).to(device)\n        with torch.no_grad():\n            desc = model.forward_features(timg.to(device)).mean(dim=(-1,2))#\n            #print (desc.shape)\n            desc = desc.view(1, -1)\n            desc_norm = F.normalize(desc, dim=1, p=2)\n        #print (desc_norm)\n        global_descs_convnext.append(desc_norm.detach().cpu())\n    global_descs_all = torch.cat(global_descs_convnext, dim=0)\n    return global_descs_all\n\n\ndef get_img_pairs_exhaustive(img_fnames):\n    index_pairs = []\n    for i in range(len(img_fnames)):\n        for j in range(i+1, len(img_fnames)):\n            index_pairs.append((i,j))\n    return index_pairs\n\n\ndef get_image_pairs_shortlist(fnames,\n                              sim_th = 0.6, # should be strict\n                              min_pairs = 20,\n                              exhaustive_if_less = 20,\n                              device=torch.device('cpu')):\n    num_imgs = len(fnames)\n\n    if num_imgs <= exhaustive_if_less:\n        return get_img_pairs_exhaustive(fnames)\n\n    model = timm.create_model('tf_efficientnet_b7',\n                              checkpoint_path='/kaggle/input/tf-efficientnet/pytorch/tf-efficientnet-b7/1/tf_efficientnet_b7_ra-6c08e654.pth')\n    model.eval()\n    descs = get_global_desc(fnames, model, device=device)\n    dm = torch.cdist(descs, descs, p=2).detach().cpu().numpy()\n    # removing half\n    mask = dm <= sim_th\n    total = 0\n    matching_list = []\n    ar = np.arange(num_imgs)\n    already_there_set = []\n    for st_idx in range(num_imgs-1):\n        mask_idx = mask[st_idx]\n        to_match = ar[mask_idx]\n        if len(to_match) < min_pairs:\n            to_match = np.argsort(dm[st_idx])[:min_pairs]  \n        for idx in to_match:\n            if st_idx == idx:\n                continue\n            if dm[st_idx, idx] < 1000:\n                matching_list.append(tuple(sorted((st_idx, idx.item()))))\n                total+=1\n    matching_list = sorted(list(set(matching_list)))\n    return matching_list","metadata":{"_kg_hide-output":true,"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def get_image_pairs_shortlist(fnames,\n                              exhaustive_if_less = 20,\n                              nneighbor=60,\n                              device=torch.device('cpu'),\n                              th=0.1):\n    num_imgs = len(fnames)\n    if num_imgs <= exhaustive_if_less or num_imgs <= nneighbor :\n        return get_img_pairs_exhaustive(fnames)\n    model = timm.create_model('tf_efficientnet_b6',\n                             checkpoint_path='/kaggle/input/tf-efficientnet/pytorch/tf-efficientnet-b6/1/tf_efficientnet_b6_aa-80ba17e4.pth')\n    model.eval()\n    model2 = timm.create_model('tf_efficientnet_b7',\n                             checkpoint_path='/kaggle/input/tf-efficientnet/pytorch/tf-efficientnet-b7/1/tf_efficientnet_b7_ra-6c08e654.pth')                  \n    model2.eval()\n    descs = get_global_desc(fnames, model, model2, device=device)\n    del model, model2\n    torch.cuda.empty_cache()\n    import gc\n    gc.collect()\n    dm = torch.einsum('bi,ki->bk', descs, descs).detach().cpu() \n    value, index = torch.topk(dm, k=nneighbor, dim=1)\n    matching_list = []\n    for i in range(num_imgs-1):\n        for t in index[i][value[i]>th]:\n            if t == i:\n                continue\n            matching_list.append(tuple(sorted((i, t.item()))))\n    matching_list = sorted(list(set(matching_list)))\n    return matching_list","metadata":{"_kg_hide-output":true,"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# COLMAP utilities","metadata":{}},{"cell_type":"code","source":"# Code to manipulate a colmap database.\n# Forked from https://github.com/colmap/colmap/blob/dev/scripts/python/database.py\n\n# Copyright (c) 2018, ETH Zurich and UNC Chapel Hill.\n# All rights reserved.\n#\n# Redistribution and use in source and binary forms, with or without\n# modification, are permitted provided that the following conditions are met:\n#\n#     * Redistributions of source code must retain the above copyright\n#       notice, this list of conditions and the following disclaimer.\n#\n#     * Redistributions in binary form must reproduce the above copyright\n#       notice, this list of conditions and the following disclaimer in the\n#       documentation and/or other materials provided with the distribution.\n#\n#     * Neither the name of ETH Zurich and UNC Chapel Hill nor the names of\n#       its contributors may be used to endorse or promote products derived\n#       from this software without specific prior written permission.\n#\n# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\"\n# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE\n# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE\n# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR CONTRIBUTORS BE\n# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR\n# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF\n# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS\n# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN\n# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)\n# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE\n# POSSIBILITY OF SUCH DAMAGE.\n#\n# Author: Johannes L. Schoenberger (jsch-at-demuc-dot-de)\n\n# This script is based on an original implementation by True Price.\n\nimport sys\nimport sqlite3\nimport numpy as np\n\n\nIS_PYTHON3 = sys.version_info[0] >= 3\n\nMAX_IMAGE_ID = 2**31 - 1\n\nCREATE_CAMERAS_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS cameras (\n    camera_id INTEGER PRIMARY KEY AUTOINCREMENT NOT NULL,\n    model INTEGER NOT NULL,\n    width INTEGER NOT NULL,\n    height INTEGER NOT NULL,\n    params BLOB,\n    prior_focal_length INTEGER NOT NULL)\"\"\"\n\nCREATE_DESCRIPTORS_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS descriptors (\n    image_id INTEGER PRIMARY KEY NOT NULL,\n    rows INTEGER NOT NULL,\n    cols INTEGER NOT NULL,\n    data BLOB,\n    FOREIGN KEY(image_id) REFERENCES images(image_id) ON DELETE CASCADE)\"\"\"\n\nCREATE_IMAGES_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS images (\n    image_id INTEGER PRIMARY KEY AUTOINCREMENT NOT NULL,\n    name TEXT NOT NULL UNIQUE,\n    camera_id INTEGER NOT NULL,\n    prior_qw REAL,\n    prior_qx REAL,\n    prior_qy REAL,\n    prior_qz REAL,\n    prior_tx REAL,\n    prior_ty REAL,\n    prior_tz REAL,\n    CONSTRAINT image_id_check CHECK(image_id >= 0 and image_id < {}),\n    FOREIGN KEY(camera_id) REFERENCES cameras(camera_id))\n\"\"\".format(MAX_IMAGE_ID)\n\nCREATE_TWO_VIEW_GEOMETRIES_TABLE = \"\"\"\nCREATE TABLE IF NOT EXISTS two_view_geometries (\n    pair_id INTEGER PRIMARY KEY NOT NULL,\n    rows INTEGER NOT NULL,\n    cols INTEGER NOT NULL,\n    data BLOB,\n    config INTEGER NOT NULL,\n    F BLOB,\n    E BLOB,\n    H BLOB)\n\"\"\"\n\nCREATE_KEYPOINTS_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS keypoints (\n    image_id INTEGER PRIMARY KEY NOT NULL,\n    rows INTEGER NOT NULL,\n    cols INTEGER NOT NULL,\n    data BLOB,\n    FOREIGN KEY(image_id) REFERENCES images(image_id) ON DELETE CASCADE)\n\"\"\"\n\nCREATE_MATCHES_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS matches (\n    pair_id INTEGER PRIMARY KEY NOT NULL,\n    rows INTEGER NOT NULL,\n    cols INTEGER NOT NULL,\n    data BLOB)\"\"\"\n\nCREATE_NAME_INDEX = \\\n    \"CREATE UNIQUE INDEX IF NOT EXISTS index_name ON images(name)\"\n\nCREATE_ALL = \"; \".join([\n    CREATE_CAMERAS_TABLE,\n    CREATE_IMAGES_TABLE,\n    CREATE_KEYPOINTS_TABLE,\n    CREATE_DESCRIPTORS_TABLE,\n    CREATE_MATCHES_TABLE,\n    CREATE_TWO_VIEW_GEOMETRIES_TABLE,\n    CREATE_NAME_INDEX\n])\n\n\ndef image_ids_to_pair_id(image_id1, image_id2):\n    if image_id1 > image_id2:\n        image_id1, image_id2 = image_id2, image_id1\n    return image_id1 * MAX_IMAGE_ID + image_id2\n\n\ndef pair_id_to_image_ids(pair_id):\n    image_id2 = pair_id % MAX_IMAGE_ID\n    image_id1 = (pair_id - image_id2) / MAX_IMAGE_ID\n    return image_id1, image_id2\n\n\ndef array_to_blob(array):\n    if IS_PYTHON3:\n        return array.tostring()\n    else:\n        return np.getbuffer(array)\n\n\ndef blob_to_array(blob, dtype, shape=(-1,)):\n    if IS_PYTHON3:\n        return np.fromstring(blob, dtype=dtype).reshape(*shape)\n    else:\n        return np.frombuffer(blob, dtype=dtype).reshape(*shape)\n\n\nclass COLMAPDatabase(sqlite3.Connection):\n\n    @staticmethod\n    def connect(database_path):\n        return sqlite3.connect(database_path, factory=COLMAPDatabase)\n\n\n    def __init__(self, *args, **kwargs):\n        super(COLMAPDatabase, self).__init__(*args, **kwargs)\n\n        self.create_tables = lambda: self.executescript(CREATE_ALL)\n        self.create_cameras_table = \\\n            lambda: self.executescript(CREATE_CAMERAS_TABLE)\n        self.create_descriptors_table = \\\n            lambda: self.executescript(CREATE_DESCRIPTORS_TABLE)\n        self.create_images_table = \\\n            lambda: self.executescript(CREATE_IMAGES_TABLE)\n        self.create_two_view_geometries_table = \\\n            lambda: self.executescript(CREATE_TWO_VIEW_GEOMETRIES_TABLE)\n        self.create_keypoints_table = \\\n            lambda: self.executescript(CREATE_KEYPOINTS_TABLE)\n        self.create_matches_table = \\\n            lambda: self.executescript(CREATE_MATCHES_TABLE)\n        self.create_name_index = lambda: self.executescript(CREATE_NAME_INDEX)\n\n    def add_camera(self, model, width, height, params,\n                   prior_focal_length=False, camera_id=None):\n        params = np.asarray(params, np.float64)\n        cursor = self.execute(\n            \"INSERT INTO cameras VALUES (?, ?, ?, ?, ?, ?)\",\n            (camera_id, model, width, height, array_to_blob(params),\n             prior_focal_length))\n        return cursor.lastrowid\n\n    def add_image(self, name, camera_id,\n                  prior_q=np.zeros(4), prior_t=np.zeros(3), image_id=None):\n        cursor = self.execute(\n            \"INSERT INTO images VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)\",\n            (image_id, name, camera_id, prior_q[0], prior_q[1], prior_q[2],\n             prior_q[3], prior_t[0], prior_t[1], prior_t[2]))\n        return cursor.lastrowid\n\n    def add_keypoints(self, image_id, keypoints):\n        assert(len(keypoints.shape) == 2)\n        assert(keypoints.shape[1] in [2, 4, 6])\n\n        keypoints = np.asarray(keypoints, np.float32)\n        self.execute(\n            \"INSERT INTO keypoints VALUES (?, ?, ?, ?)\",\n            (image_id,) + keypoints.shape + (array_to_blob(keypoints),))\n\n    def add_descriptors(self, image_id, descriptors):\n        descriptors = np.ascontiguousarray(descriptors, np.uint8)\n        self.execute(\n            \"INSERT INTO descriptors VALUES (?, ?, ?, ?)\",\n            (image_id,) + descriptors.shape + (array_to_blob(descriptors),))\n\n    def add_matches(self, image_id1, image_id2, matches):\n        assert(len(matches.shape) == 2)\n        assert(matches.shape[1] == 2)\n\n        if image_id1 > image_id2:\n            matches = matches[:,::-1]\n\n        pair_id = image_ids_to_pair_id(image_id1, image_id2)\n        matches = np.asarray(matches, np.uint32)\n        self.execute(\n            \"INSERT INTO matches VALUES (?, ?, ?, ?)\",\n            (pair_id,) + matches.shape + (array_to_blob(matches),))\n\n    def add_two_view_geometry(self, image_id1, image_id2, matches,\n                              F=np.eye(3), E=np.eye(3), H=np.eye(3), config=2):\n        assert(len(matches.shape) == 2)\n        assert(matches.shape[1] == 2)\n\n        if image_id1 > image_id2:\n            matches = matches[:,::-1]\n\n        pair_id = image_ids_to_pair_id(image_id1, image_id2)\n        matches = np.asarray(matches, np.uint32)\n        F = np.asarray(F, dtype=np.float64)\n        E = np.asarray(E, dtype=np.float64)\n        H = np.asarray(H, dtype=np.float64)\n        self.execute(\n            \"INSERT INTO two_view_geometries VALUES (?, ?, ?, ?, ?, ?, ?, ?)\",\n            (pair_id,) + matches.shape + (array_to_blob(matches), config,\n             array_to_blob(F), array_to_blob(E), array_to_blob(H)))","metadata":{"_kg_hide-output":true,"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# h5 to colmap db","metadata":{}},{"cell_type":"code","source":"# Code to interface DISK with Colmap.\n# Forked from https://github.com/cvlab-epfl/disk/blob/37f1f7e971cea3055bb5ccfc4cf28bfd643fa339/colmap/h5_to_db.py\n\n#  Copyright [2020] [Michał Tyszkiewicz, Pascal Fua, Eduard Trulls]\n#\n#   Licensed under the Apache License, Version 2.0 (the \"License\");\n#   you may not use this file except in compliance with the License.\n#   You may obtain a copy of the License at\n#\n#       http://www.apache.org/licenses/LICENSE-2.0\n#\n#   Unless required by applicable law or agreed to in writing, software\n#   distributed under the License is distributed on an \"AS IS\" BASIS,\n#   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n#   See the License for the specific language governing permissions and\n#   limitations under the License.\n\nimport os, argparse, h5py, warnings\nimport numpy as np\nfrom tqdm import tqdm\nfrom PIL import Image, ExifTags\n\n\ndef get_focal(image_path, err_on_default=False):\n    image         = Image.open(image_path)\n    max_size      = max(image.size)\n\n    exif = image.getexif()\n    focal = None\n    if exif is not None:\n        focal_35mm = None\n        # https://github.com/colmap/colmap/blob/d3a29e203ab69e91eda938d6e56e1c7339d62a99/src/util/bitmap.cc#L299\n        for tag, value in exif.items():\n            focal_35mm = None\n            if ExifTags.TAGS.get(tag, None) == 'FocalLengthIn35mmFilm':\n                focal_35mm = float(value)\n                break\n\n        if focal_35mm is not None:\n            focal = focal_35mm / 35. * max_size\n    \n    if focal is None:\n        if err_on_default:\n            raise RuntimeError(\"Failed to find focal length\")\n\n        # failed to find it in exif, use prior\n        FOCAL_PRIOR = 1.2\n        focal = FOCAL_PRIOR * max_size\n\n    return focal\n\ndef create_camera(db, image_path, camera_model):\n    image         = Image.open(image_path)\n    width, height = image.size\n\n    focal = get_focal(image_path)\n\n    if camera_model == 'simple-pinhole':\n        model = 0 # simple pinhole\n        param_arr = np.array([focal, width / 2, height / 2])\n    if camera_model == 'pinhole':\n        model = 1 # pinhole\n        param_arr = np.array([focal, focal, width / 2, height / 2])\n    elif camera_model == 'simple-radial':\n        model = 2 # simple radial\n        param_arr = np.array([focal, width / 2, height / 2, 0.1])\n    elif camera_model == 'opencv':\n        model = 4 # opencv\n        param_arr = np.array([focal, focal, width / 2, height / 2, 0., 0., 0., 0.])\n         \n    return db.add_camera(model, width, height, param_arr)\n\n\ndef add_keypoints(db, h5_path, image_path, img_ext, camera_model, single_camera = True):\n    keypoint_f = h5py.File(os.path.join(h5_path, 'keypoints.h5'), 'r')\n\n    camera_id = None\n    fname_to_id = {}\n    for filename in tqdm(list(keypoint_f.keys())):\n        keypoints = keypoint_f[filename][()]\n\n        fname_with_ext = filename# + img_ext\n        path = os.path.join(image_path, fname_with_ext)\n        if not os.path.isfile(path):\n            raise IOError(f'Invalid image path {path}')\n\n        if camera_id is None or not single_camera:\n            camera_id = create_camera(db, path, camera_model)\n        image_id = db.add_image(fname_with_ext, camera_id)\n        fname_to_id[filename] = image_id\n\n        db.add_keypoints(image_id, keypoints)\n\n    return fname_to_id\n\ndef add_matches(db, h5_path, fname_to_id):\n    match_file = h5py.File(os.path.join(h5_path, 'matches.h5'), 'r')\n    \n    added = set()\n    n_keys = len(match_file.keys())\n    n_total = (n_keys * (n_keys - 1)) // 2\n\n    with tqdm(total=n_total) as pbar:\n        for key_1 in match_file.keys():\n            group = match_file[key_1]\n            for key_2 in group.keys():\n                id_1 = fname_to_id[key_1]\n                id_2 = fname_to_id[key_2]\n\n                pair_id = image_ids_to_pair_id(id_1, id_2)\n                if pair_id in added:\n                    warnings.warn(f'Pair {pair_id} ({id_1}, {id_2}) already added!')\n                    continue\n            \n                matches = group[key_2][()]\n                db.add_matches(id_1, id_2, matches)\n\n                added.add(pair_id)\n\n                pbar.update(1)","metadata":{"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Making kornia local features loading w/o internet\nclass KeyNetAffNetHardNet(KF.LocalFeature):\n    \"\"\"Convenience module, which implements KeyNet detector + AffNet + HardNet descriptor.\n\n    .. image:: _static/img/keynet_affnet.jpg\n    \"\"\"\n\n    def __init__(\n        self,\n        num_features: int = 5000,\n        upright: bool = False,\n        device = torch.device('cpu'),\n        scale_laf: float = 1.0,\n    ):\n        ori_module = KF.PassLAF() if upright else KF.LAFOrienter(angle_detector=KF.OriNet(False)).eval()\n        if not upright:\n            weights = torch.load('/kaggle/input/kornia-local-feature-weights/OriNet.pth')['state_dict']\n            ori_module.angle_detector.load_state_dict(weights)\n        detector = KF.KeyNetDetector(\n            False, num_features=num_features, ori_module=ori_module, aff_module=KF.LAFAffNetShapeEstimator(False).eval()\n        ).to(device)\n        kn_weights = torch.load('/kaggle/input/kornia-local-feature-weights/keynet_pytorch.pth')['state_dict']\n        detector.model.load_state_dict(kn_weights)\n        affnet_weights = torch.load('/kaggle/input/kornia-local-feature-weights/AffNet.pth')['state_dict']\n        detector.aff.load_state_dict(affnet_weights)\n        \n        hardnet = KF.HardNet(False).eval()\n        hn_weights = torch.load('/kaggle/input/kornia-local-feature-weights/HardNetLib.pth')['state_dict']\n        hardnet.load_state_dict(hn_weights)\n        descriptor = KF.LAFDescriptor(hardnet, patch_size=32, grayscale_descriptor=True).to(device)\n        super().__init__(detector, descriptor, scale_laf)","metadata":{"_kg_hide-output":true,"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def detect_features(img_fnames,\n                    num_feats = 2048,\n                    upright = False,\n                    device=torch.device('cpu'),\n                    feature_dir = '.featureout',\n                    resize_small_edge_to = 600):\n    if LOCAL_FEATURE == 'DISK':\n        # Load DISK from Kaggle models so it can run when the notebook is offline.\n        disk = KF.DISK().to(device)\n        pretrained_dict = torch.load('/kaggle/input/disk/pytorch/depth-supervision/1/loftr_outdoor.ckpt', map_location=device)\n        disk.load_state_dict(pretrained_dict['extractor'])\n        disk.eval()\n    if LOCAL_FEATURE == 'KeyNetAffNetHardNet':\n        feature = KeyNetAffNetHardNet(num_feats, upright, device).to(device).eval()\n    if not os.path.isdir(feature_dir):\n        os.makedirs(feature_dir)\n    with h5py.File(f'{feature_dir}/lafs.h5', mode='w') as f_laf, \\\n         h5py.File(f'{feature_dir}/keypoints.h5', mode='w') as f_kp, \\\n         h5py.File(f'{feature_dir}/descriptors.h5', mode='w') as f_desc:\n        for img_path in progress_bar(img_fnames):\n            img_fname = img_path.split('/')[-1]\n            key = img_fname\n            with torch.inference_mode():\n                timg = load_torch_image(img_path, device=device)\n                H, W = timg.shape[2:]\n                if resize_small_edge_to is None:\n                    timg_resized = timg\n                else:\n                    timg_resized = K.geometry.resize(timg, resize_small_edge_to, antialias=True)\n                    print(f'Resized {timg.shape} to {timg_resized.shape} (resize_small_edge_to={resize_small_edge_to})')\n                h, w = timg_resized.shape[2:]\n                if LOCAL_FEATURE == 'DISK':\n                    features = disk(timg_resized, num_feats, pad_if_not_divisible=True)[0]\n                    kps1, descs = features.keypoints, features.descriptors\n                    \n                    lafs = KF.laf_from_center_scale_ori(kps1[None], torch.ones(1, len(kps1), 1, 1, device=device))\n                if LOCAL_FEATURE == 'KeyNetAffNetHardNet':\n                    lafs, resps, descs = feature(K.color.rgb_to_grayscale(timg_resized))\n                lafs[:,:,0,:] *= float(W) / float(w)\n                lafs[:,:,1,:] *= float(H) / float(h)\n                desc_dim = descs.shape[-1]\n                kpts = KF.get_laf_center(lafs).reshape(-1, 2).detach().cpu().numpy()\n                descs = descs.reshape(-1, desc_dim).detach().cpu().numpy()\n                f_laf[key] = lafs.detach().cpu().numpy()\n                f_kp[key] = kpts\n                f_desc[key] = descs\n    return\n\ndef get_unique_idxs(A, dim=0):\n    # https://stackoverflow.com/questions/72001505/how-to-get-unique-elements-and-their-firstly-appeared-indices-of-a-pytorch-tenso\n    unique, idx, counts = torch.unique(A, dim=dim, sorted=True, return_inverse=True, return_counts=True)\n    _, ind_sorted = torch.sort(idx, stable=True)\n    cum_sum = counts.cumsum(0)\n    cum_sum = torch.cat((torch.tensor([0],device=cum_sum.device), cum_sum[:-1]))\n    first_indices = ind_sorted[cum_sum]\n    return first_indices\n\ndef match_features(img_fnames,\n                   index_pairs,\n                   feature_dir = '.featureout',\n                   device=torch.device('cpu'),\n                   min_matches=15, \n                   force_mutual = True,\n                   matching_alg='smnn'\n                  ):\n    assert matching_alg in ['smnn', 'adalam']\n    with h5py.File(f'{feature_dir}/lafs.h5', mode='r') as f_laf, \\\n         h5py.File(f'{feature_dir}/descriptors.h5', mode='r') as f_desc, \\\n        h5py.File(f'{feature_dir}/matches.h5', mode='w') as f_match:\n\n        for pair_idx in progress_bar(index_pairs):\n                    idx1, idx2 = pair_idx\n                    fname1, fname2 = img_fnames[idx1], img_fnames[idx2]\n                    key1, key2 = fname1.split('/')[-1], fname2.split('/')[-1]\n                    lafs1 = torch.from_numpy(f_laf[key1][...]).to(device)\n                    lafs2 = torch.from_numpy(f_laf[key2][...]).to(device)\n                    desc1 = torch.from_numpy(f_desc[key1][...]).to(device)\n                    desc2 = torch.from_numpy(f_desc[key2][...]).to(device)\n                    if matching_alg == 'adalam':\n                        img1, img2 = cv2.imread(fname1), cv2.imread(fname2)\n                        hw1, hw2 = img1.shape[:2], img2.shape[:2]\n                        adalam_config = KF.adalam.get_adalam_default_config()\n                        #adalam_config['orientation_difference_threshold'] = None\n                        #adalam_config['scale_rate_threshold'] = None\n                        adalam_config['force_seed_mnn']= False\n                        adalam_config['search_expansion'] = 16\n                        adalam_config['ransac_iters'] = 128\n                        adalam_config['device'] = device\n                        dists, idxs = KF.match_adalam(desc1, desc2,\n                                                      lafs1, lafs2, # Adalam takes into account also geometric information\n                                                      hw1=hw1, hw2=hw2,\n                                                      config=adalam_config) # Adalam also benefits from knowing image size\n                    else:\n                        dists, idxs = KF.match_smnn(desc1, desc2, 0.98)\n                    if len(idxs)  == 0:\n                        continue\n                    # Force mutual nearest neighbors\n                    if force_mutual:\n                        first_indices = get_unique_idxs(idxs[:,1])\n                        idxs = idxs[first_indices]\n                        dists = dists[first_indices]\n                    n_matches = len(idxs)\n                    if False:\n                        print (f'{key1}-{key2}: {n_matches} matches')\n                    group  = f_match.require_group(key1)\n                    if n_matches >= min_matches:\n                         group.create_dataset(key2, data=idxs.detach().cpu().numpy().reshape(-1, 2))\n    return\n\ndef match_loftr(img_fnames,\n                   index_pairs,\n                   feature_dir = '.featureout_loftr',\n                   device=torch.device('cpu'),\n                   min_matches=15, resize_to_ = (640, 480)):\n    matcher = KF.LoFTR(pretrained=None)\n    matcher.load_state_dict(torch.load('/kaggle/input/loftr/pytorch/outdoor/1/loftr_outdoor.ckpt')['state_dict'])\n    matcher = matcher.to(device).eval()\n\n    # First we do pairwise matching, and then extract \"keypoints\" from loftr matches.\n    with h5py.File(f'{feature_dir}/matches_loftr.h5', mode='w') as f_match:\n        for pair_idx in progress_bar(index_pairs):\n            idx1, idx2 = pair_idx\n            fname1, fname2 = img_fnames[idx1], img_fnames[idx2]\n            key1, key2 = fname1.split('/')[-1], fname2.split('/')[-1]\n            # Load img1\n            timg1 = K.color.rgb_to_grayscale(load_torch_image(fname1, device=device))\n            H1, W1 = timg1.shape[2:]\n            if H1 < W1:\n                resize_to = resize_to_[1], resize_to_[0]\n            else:\n                resize_to = resize_to_\n            timg_resized1 = K.geometry.resize(timg1, resize_to, antialias=True)\n            h1, w1 = timg_resized1.shape[2:]\n\n            # Load img2\n            timg2 = K.color.rgb_to_grayscale(load_torch_image(fname2, device=device))\n            H2, W2 = timg2.shape[2:]\n            if H2 < W2:\n                resize_to2 = resize_to[1], resize_to[0]\n            else:\n                resize_to2 = resize_to_\n            timg_resized2 = K.geometry.resize(timg2, resize_to2, antialias=True)\n            h2, w2 = timg_resized2.shape[2:]\n            with torch.inference_mode():\n                input_dict = {\"image0\": timg_resized1,\"image1\": timg_resized2}\n                correspondences = matcher(input_dict)\n            mkpts0 = correspondences['keypoints0'].cpu().numpy()\n            mkpts1 = correspondences['keypoints1'].cpu().numpy()\n\n            mkpts0[:,0] *= float(W1) / float(w1)\n            mkpts0[:,1] *= float(H1) / float(h1)\n\n            mkpts1[:,0] *= float(W2) / float(w2)\n            mkpts1[:,1] *= float(H2) / float(h2)\n\n            n_matches = len(mkpts1)\n            group  = f_match.require_group(key1)\n            if n_matches >= min_matches:\n                 group.create_dataset(key2, data=np.concatenate([mkpts0, mkpts1], axis=1))\n\n    # Let's find unique loftr pixels and group them together.\n    kpts = defaultdict(list)\n    match_indexes = defaultdict(dict)\n    total_kpts=defaultdict(int)\n    with h5py.File(f'{feature_dir}/matches_loftr.h5', mode='r') as f_match:\n        for k1 in f_match.keys():\n            group  = f_match[k1]\n            for k2 in group.keys():\n                matches = group[k2][...]\n                total_kpts[k1]\n                kpts[k1].append(matches[:, :2])\n                kpts[k2].append(matches[:, 2:])\n                current_match = torch.arange(len(matches)).reshape(-1, 1).repeat(1, 2)\n                current_match[:, 0]+=total_kpts[k1]\n                current_match[:, 1]+=total_kpts[k2]\n                total_kpts[k1]+=len(matches)\n                total_kpts[k2]+=len(matches)\n                match_indexes[k1][k2]=current_match\n\n    for k in kpts.keys():\n        kpts[k] = np.round(np.concatenate(kpts[k], axis=0))\n    unique_kpts = {}\n    unique_match_idxs = {}\n    out_match = defaultdict(dict)\n    for k in kpts.keys():\n        uniq_kps, uniq_reverse_idxs = torch.unique(torch.from_numpy(kpts[k]),dim=0, return_inverse=True)\n        unique_match_idxs[k] = uniq_reverse_idxs\n        unique_kpts[k] = uniq_kps.numpy()\n    for k1, group in match_indexes.items():\n        for k2, m in group.items():\n            m2 = deepcopy(m)\n            m2[:,0] = unique_match_idxs[k1][m2[:,0]]\n            m2[:,1] = unique_match_idxs[k2][m2[:,1]]\n            mkpts = np.concatenate([unique_kpts[k1][ m2[:,0]],\n                                    unique_kpts[k2][  m2[:,1]],\n                                   ],\n                                   axis=1)\n            unique_idxs_current = get_unique_idxs(torch.from_numpy(mkpts), dim=0)\n            m2_semiclean = m2[unique_idxs_current]\n            unique_idxs_current1 = get_unique_idxs(m2_semiclean[:, 0], dim=0)\n            m2_semiclean = m2_semiclean[unique_idxs_current1]\n            unique_idxs_current2 = get_unique_idxs(m2_semiclean[:, 1], dim=0)\n            m2_semiclean2 = m2_semiclean[unique_idxs_current2]\n            out_match[k1][k2] = m2_semiclean2.numpy()\n    with h5py.File(f'{feature_dir}/keypoints.h5', mode='w') as f_kp:\n        for k, kpts1 in unique_kpts.items():\n            f_kp[k] = kpts1\n    \n    with h5py.File(f'{feature_dir}/matches.h5', mode='w') as f_match:\n        for k1, gr in out_match.items():\n            group  = f_match.require_group(k1)\n            for k2, match in gr.items():\n                group[k2] = match\n    return\n\ndef import_into_colmap(img_dir,\n                       feature_dir ='.featureout',\n                       database_path = 'colmap.db',\n                       img_ext='.jpg'):\n    db = COLMAPDatabase.connect(database_path)\n    db.create_tables()\n    single_camera = False\n    fname_to_id = add_keypoints(db, feature_dir, img_dir, img_ext, 'simple-radial', single_camera)\n    add_matches(\n        db,\n        feature_dir,\n        fname_to_id,\n    )\n\n    db.commit()\n    return","metadata":{"_kg_hide-output":true,"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"src = '/kaggle/input/image-matching-challenge-2023'","metadata":{"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Get data from csv.\n\ndata_dict = {}\nwith open(f'{src}/sample_submission.csv', 'r') as f:\n    for i, l in enumerate(f):\n        # Skip header.\n        if l and i > 0:\n            image, dataset, scene, _, _ = l.strip().split(',')\n            if dataset not in data_dict:\n                data_dict[dataset] = {}\n            if scene not in data_dict[dataset]:\n                data_dict[dataset][scene] = []\n            data_dict[dataset][scene].append(image)","metadata":{"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"for dataset in data_dict:\n    for scene in data_dict[dataset]:\n        print(f'{dataset} / {scene} -> {len(data_dict[dataset][scene])} images')","metadata":{"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"out_results = {}\ntimings = {\"shortlisting\":[],\n           \"feature_detection\": [],\n           \"feature_matching\":[],\n           \"RANSAC\": [],\n           \"Reconstruction\": []}","metadata":{"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Function to create a submission file.\ndef create_submission(out_results, data_dict):\n    with open(f'submission.csv', 'w') as f:\n        f.write('image_path,dataset,scene,rotation_matrix,translation_vector\\n')\n        for dataset in data_dict:\n            if dataset in out_results:\n                res = out_results[dataset]\n            else:\n                res = {}\n            for scene in data_dict[dataset]:\n                if scene in res:\n                    scene_res = res[scene]\n                else:\n                    scene_res = {\"R\":{}, \"t\":{}}\n                for image in data_dict[dataset][scene]:\n                    if image in scene_res:\n                        print (image)\n                        R = scene_res[image]['R'].reshape(-1)\n                        T = scene_res[image]['t'].reshape(-1)\n                    else:\n                        R = np.eye(3).reshape(-1)\n                        T = np.zeros((3))\n                    f.write(f'{image},{dataset},{scene},{arr_to_str(R)},{arr_to_str(T)}\\n')","metadata":{"_kg_hide-input":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"gc.collect()\ndatasets = []\nfor dataset in data_dict:\n    datasets.append(dataset)\n\nfor dataset in datasets:\n    print(dataset)\n    if dataset not in out_results:\n        out_results[dataset] = {}\n    for scene in data_dict[dataset]:\n        print(scene)\n        # Fail gently if the notebook has not been submitted and the test data is not populated.\n        # You may want to run this on the training data in that case?\n        img_dir = f'{src}/test/{dataset}/{scene}/images'\n        if not os.path.exists(img_dir):\n            continue\n        # Wrap the meaty part in a try-except block.\n        try:\n            out_results[dataset][scene] = {}\n            img_fnames = [f'{src}/test/{x}' for x in data_dict[dataset][scene]]\n            print (f\"Got {len(img_fnames)} images\")\n            feature_dir = f'featureout/{dataset}_{scene}'\n            if not os.path.isdir(feature_dir):\n                os.makedirs(feature_dir, exist_ok=True)\n            t=time()\n            index_pairs = get_image_pairs_shortlist(img_fnames,\n                                  sim_th = 0.5, # should be strict\n                                  min_pairs = 20, # we select at least min_pairs PER IMAGE with biggest similarity\n                                  exhaustive_if_less = 20,\n                                  device=device)\n            t=time() -t \n            timings['shortlisting'].append(t)\n            print (f'{len(index_pairs)}, pairs to match, {t:.4f} sec')\n            gc.collect()\n            t=time()\n            if LOCAL_FEATURE != 'LoFTR':\n                detect_features(img_fnames, \n                                2048,\n                                feature_dir=feature_dir,\n                                upright=True,\n                                device=device,\n                                resize_small_edge_to=600\n                               )\n                gc.collect()\n                t=time() -t \n                timings['feature_detection'].append(t)\n                print(f'Features detected in  {t:.4f} sec')\n                t=time()\n                match_features(img_fnames, index_pairs, feature_dir=feature_dir,device=device)\n            else:\n                match_loftr(img_fnames, index_pairs, feature_dir=feature_dir, device=device, resize_to_=(600, 800))\n            t=time() -t \n            timings['feature_matching'].append(t)\n            print(f'Features matched in  {t:.4f} sec')\n            database_path = f'{feature_dir}/colmap.db'\n            if os.path.isfile(database_path):\n                os.remove(database_path)\n            gc.collect()\n            import_into_colmap(img_dir, feature_dir=feature_dir,database_path=database_path)\n            output_path = f'{feature_dir}/colmap_rec_{LOCAL_FEATURE}'\n\n            t=time()\n            pycolmap.match_exhaustive(database_path)\n            t=time() - t \n            timings['RANSAC'].append(t)\n            print(f'RANSAC in  {t:.4f} sec')\n\n            t=time()\n            # By default colmap does not generate a reconstruction if less than 10 images are registered. Lower it to 3.\n            mapper_options = pycolmap.IncrementalMapperOptions()\n            mapper_options.min_model_size = 3\n            os.makedirs(output_path, exist_ok=True)\n            maps = pycolmap.incremental_mapping(database_path=database_path, image_path=img_dir, output_path=output_path, options=mapper_options)\n            print(maps)\n            #clear_output(wait=False)\n            t=time() - t\n            timings['Reconstruction'].append(t)\n            print(f'Reconstruction done in  {t:.4f} sec')\n            imgs_registered  = 0\n            best_idx = None\n            print (\"Looking for the best reconstruction\")\n            if isinstance(maps, dict):\n                for idx1, rec in maps.items():\n                    print (idx1, rec.summary())\n                    if len(rec.images) > imgs_registered:\n                        imgs_registered = len(rec.images)\n                        best_idx = idx1\n            if best_idx is not None:\n                print (maps[best_idx].summary())\n                for k, im in maps[best_idx].images.items():\n                    key1 = f'{dataset}/{scene}/images/{im.name}'\n                    out_results[dataset][scene][key1] = {}\n                    out_results[dataset][scene][key1][\"R\"] = deepcopy(im.rotmat())\n                    out_results[dataset][scene][key1][\"t\"] = deepcopy(np.array(im.tvec))\n            print(f'Registered: {dataset} / {scene} -> {len(out_results[dataset][scene])} images')\n            print(f'Total: {dataset} / {scene} -> {len(data_dict[dataset][scene])} images')\n            create_submission(out_results, data_dict)\n            gc.collect()\n        except:\n            pass","metadata":{"_kg_hide-input":false},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"create_submission(out_results, data_dict)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!python -m pip install --no-deps /kaggle/input/dependencies-imc/pycolmap/pycolmap-0.4.0-cp310-cp310-manylinux2014_x86_64.whl\n!python -m pip install --no-deps /kaggle/input/dependencies-imc/safetensors/safetensors-0.4.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl\n!python -m pip install --no-index --find-links=/kaggle/input/dependencies-imc/transformers/ transformers > /dev/null\n!python -m pip install  --no-deps /kaggle/input/imc2024-packages-lightglue-rerun-kornia/lightglue-0.0-py3-none-any.whl\n\n# dkm\n!python -m pip install --no-index --find-links=/kaggle/input/dkm-dependencies/packages einops > /dev/null\n\n# match former\n!python -m pip install --no-index --find-links=/kaggle/input/matchformer-dependencies yacs > /dev/null\n\n# lightglue models\n!mkdir -p /root/.cache/torch/hub/checkpoints\n!cp /kaggle/input/aliked/pytorch/aliked-n16/1/* /root/.cache/torch/hub/checkpoints/\n!cp /kaggle/input/lightglue/pytorch/aliked/1/* /root/.cache/torch/hub/checkpoints/\n!cp /kaggle/input/lightglue/pytorch/aliked/1/aliked_lightglue.pth /root/.cache/torch/hub/checkpoints/aliked_lightglue_v0-1_arxiv-pth\n!cp /kaggle/input/pytorch-lightglue-models/* /root/.cache/torch/hub/checkpoints/\n\n# dkm model\n!mkdir -p /root/.cache/torch/hub/checkpoints\n!cp /kaggle/input/dkm-dependencies/DKMv3_outdoor.pth /root/.cache/torch/hub/checkpoints/\n\n# check rotation\n!python -m pip install --no-index --find-links=/kaggle/input/pkg-check-orientation/ check_orientation==0.0.5 > /dev/null\n!cp /kaggle/input/pkg-check-orientation/2020-11-16_resnext50_32x4d.zip /root/.cache/torch/hub/checkpoints/\n\n%matplotlib inline","metadata":{"_kg_hide-input":true,"_kg_hide-output":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import math\nimport os\nfrom tqdm import tqdm\nfrom time import time\nfrom fastprogress import progress_bar\nimport gc\nimport numpy as np\nimport pandas as pd\nimport h5py\nfrom IPython.display import clear_output\nfrom collections import defaultdict\nfrom copy import deepcopy\nimport concurrent.futures\nfrom collections import Counter\n\n# CV/ML\nimport cv2\nimport torch\nimport torch.nn.functional as F\nfrom torch.utils.data import Dataset, DataLoader\nimport kornia as K\nimport kornia.feature as KF\nfrom PIL import Image\nimport timm\nfrom timm.data import resolve_data_config\nfrom timm.data.transforms_factory import create_transform\n\nimport torchvision\n\n# 3D reconstruction\nimport pycolmap\n\nimport glob\nimport matplotlib\nfrom matplotlib import pyplot as plt\n\n# dkm\nimport sys\nsys.path.append('/kaggle/input/dkm-dependencies/DKM/')\nfrom dkm.utils.utils import tensor_to_pil, get_tuple_transform_ops\nfrom dkm import DKMv3_outdoor\n\n# LoFTR\nfrom kornia.feature import LoFTR\n\n# LightGlue\nfrom lightglue import match_pair\nfrom lightglue import ALIKED, SuperPoint, DoGHardNet, LightGlue, DISK, SIFT\nfrom lightglue.utils import load_image, rbd\nclass CONFIG:\n    # DEBUG Settings\n    DRY_RUN = False\n    DRY_RUN_MAX_IMAGES = 10\n\n    # Pipeline settings\n    NUM_CORES = 2\n    \n    # COLMAP Reconstruction\n    CAMERA_MODEL = \"simple-radial\"\n    \n    # Rotation correction\n    ROTATION_CORRECTION = 0\n    \n    # Keypoints handling\n    MERGE_PARAMS = {\n        \"min_matches\" : 15,\n        \"filter_FundamentalMatrix\" : False,\n        \"filter_iterations\" : 10,\n        \"filter_threshold\" : 8,\n    }\n    \n    # Keypoints Extraction\n    use_aliked_lightglue = True\n    use_doghardnet_lightglue = False\n    use_superpoint_lightglue = False\n    use_disk_lightglue = False\n    use_sift_lightglue = False\n    use_loftr = False\n    use_dkm = False\n    use_superglue = False\n    use_matchformer = False\n        \n    # Keypoints Extraction Parameters\n    params_aliked_lightglue = {\n        \"num_features\" : 8192,\n        \"detection_threshold\" : 0.001,\n        \"min_matches\" : 60,\n        \"resize_to\" : 1024,\n    }\n    \n    params_doghardnet_lightglue = {\n        \"num_features\" : 8192,\n        \"detection_threshold\" : 0.005,\n        \"min_matches\" : 15,\n        \"resize_to\" : 1024,\n    }\n    \n    params_superpoint_lightglue = {\n        \"num_features\" : 8192,\n        \"detection_threshold\" : 0.005,\n        \"min_matches\" : 60,\n        \"resize_to\" : 1024,\n    }\n    \n    \n    params_sg3 = {\n        \"sg_config\" : \n        {\n            \"superpoint\": {\n                \"nms_radius\": 4, \n                \"keypoint_threshold\": 0.005,\n                \"max_keypoints\": -1,\n            },\n            \"superglue\": {\n                \"weights\": \"outdoor\",\n                \"sinkhorn_iterations\": 50,\n                \"match_threshold\": 0.3,\n            },\n        },\n        \"resize_to\": 1680,\n        \"min_matches\": 100,\n    }\n    params_sgs = [params_sg3,]\n\n\ndevice=torch.device('cuda')\n\nimport sys\nimport sqlite3\nimport numpy as np\n\n\nIS_PYTHON3 = sys.version_info[0] >= 3\n\nMAX_IMAGE_ID = 2**31 - 1\n\nCREATE_CAMERAS_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS cameras (\n    camera_id INTEGER PRIMARY KEY AUTOINCREMENT NOT NULL,\n    model INTEGER NOT NULL,\n    width INTEGER NOT NULL,\n    height INTEGER NOT NULL,\n    params BLOB,\n    prior_focal_length INTEGER NOT NULL)\"\"\"\n\nCREATE_DESCRIPTORS_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS descriptors (\n    image_id INTEGER PRIMARY KEY NOT NULL,\n    rows INTEGER NOT NULL,\n    cols INTEGER NOT NULL,\n    data BLOB,\n    FOREIGN KEY(image_id) REFERENCES images(image_id) ON DELETE CASCADE)\"\"\"\n\nCREATE_IMAGES_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS images (\n    image_id INTEGER PRIMARY KEY AUTOINCREMENT NOT NULL,\n    name TEXT NOT NULL UNIQUE,\n    camera_id INTEGER NOT NULL,\n    prior_qw REAL,\n    prior_qx REAL,\n    prior_qy REAL,\n    prior_qz REAL,\n    prior_tx REAL,\n    prior_ty REAL,\n    prior_tz REAL,\n    CONSTRAINT image_id_check CHECK(image_id >= 0 and image_id < {}),\n    FOREIGN KEY(camera_id) REFERENCES cameras(camera_id))\n\"\"\".format(MAX_IMAGE_ID)\n\nCREATE_TWO_VIEW_GEOMETRIES_TABLE = \"\"\"\nCREATE TABLE IF NOT EXISTS two_view_geometries (\n    pair_id INTEGER PRIMARY KEY NOT NULL,\n    rows INTEGER NOT NULL,\n    cols INTEGER NOT NULL,\n    data BLOB,\n    config INTEGER NOT NULL,\n    F BLOB,\n    E BLOB,\n    H BLOB)\n\"\"\"\n\nCREATE_KEYPOINTS_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS keypoints (\n    image_id INTEGER PRIMARY KEY NOT NULL,\n    rows INTEGER NOT NULL,\n    cols INTEGER NOT NULL,\n    data BLOB,\n    FOREIGN KEY(image_id) REFERENCES images(image_id) ON DELETE CASCADE)\n\"\"\"\n\nCREATE_MATCHES_TABLE = \"\"\"CREATE TABLE IF NOT EXISTS matches (\n    pair_id INTEGER PRIMARY KEY NOT NULL,\n    rows INTEGER NOT NULL,\n    cols INTEGER NOT NULL,\n    data BLOB)\"\"\"\n\nCREATE_NAME_INDEX = \\\n    \"CREATE UNIQUE INDEX IF NOT EXISTS index_name ON images(name)\"\n\nCREATE_ALL = \"; \".join([\n    CREATE_CAMERAS_TABLE,\n    CREATE_IMAGES_TABLE,\n    CREATE_KEYPOINTS_TABLE,\n    CREATE_DESCRIPTORS_TABLE,\n    CREATE_MATCHES_TABLE,\n    CREATE_TWO_VIEW_GEOMETRIES_TABLE,\n    CREATE_NAME_INDEX\n])\n\n\ndef image_ids_to_pair_id(image_id1, image_id2):\n    if image_id1 > image_id2:\n        image_id1, image_id2 = image_id2, image_id1\n    return image_id1 * MAX_IMAGE_ID + image_id2\n\n\ndef pair_id_to_image_ids(pair_id):\n    image_id2 = pair_id % MAX_IMAGE_ID\n    image_id1 = (pair_id - image_id2) / MAX_IMAGE_ID\n    return image_id1, image_id2\n\n\ndef array_to_blob(array):\n    if IS_PYTHON3:\n        return array.tostring()\n    else:\n        return np.getbuffer(array)\n\n\ndef blob_to_array(blob, dtype, shape=(-1,)):\n    if IS_PYTHON3:\n        return np.fromstring(blob, dtype=dtype).reshape(*shape)\n    else:\n        return np.frombuffer(blob, dtype=dtype).reshape(*shape)\n\n\nclass COLMAPDatabase(sqlite3.Connection):\n\n    @staticmethod\n    def connect(database_path):\n        return sqlite3.connect(database_path, factory=COLMAPDatabase)\n\n\n    def __init__(self, *args, **kwargs):\n        super(COLMAPDatabase, self).__init__(*args, **kwargs)\n\n        self.create_tables = lambda: self.executescript(CREATE_ALL)\n        self.create_cameras_table = \\\n            lambda: self.executescript(CREATE_CAMERAS_TABLE)\n        self.create_descriptors_table = \\\n            lambda: self.executescript(CREATE_DESCRIPTORS_TABLE)\n        self.create_images_table = \\\n            lambda: self.executescript(CREATE_IMAGES_TABLE)\n        self.create_two_view_geometries_table = \\\n            lambda: self.executescript(CREATE_TWO_VIEW_GEOMETRIES_TABLE)\n        self.create_keypoints_table = \\\n            lambda: self.executescript(CREATE_KEYPOINTS_TABLE)\n        self.create_matches_table = \\\n            lambda: self.executescript(CREATE_MATCHES_TABLE)\n        self.create_name_index = lambda: self.executescript(CREATE_NAME_INDEX)\n\n    def add_camera(self, model, width, height, params,\n                prior_focal_length=False, camera_id=None):\n        params = np.asarray(params, np.float64)\n        cursor = self.execute(\n            \"INSERT INTO cameras VALUES (?, ?, ?, ?, ?, ?)\",\n            (camera_id, model, width, height, array_to_blob(params),\n            prior_focal_length))\n        return cursor.lastrowid\n\n    def add_image(self, name, camera_id,\n                prior_q=np.zeros(4), prior_t=np.zeros(3), image_id=None):\n        cursor = self.execute(\n            \"INSERT INTO images VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)\",\n            (image_id, name, camera_id, prior_q[0], prior_q[1], prior_q[2],\n            prior_q[3], prior_t[0], prior_t[1], prior_t[2]))\n        return cursor.lastrowid\n\n    def add_keypoints(self, image_id, keypoints):\n        assert(len(keypoints.shape) == 2)\n        assert(keypoints.shape[1] in [2, 4, 6])\n\n        keypoints = np.asarray(keypoints, np.float32)\n        self.execute(\n            \"INSERT INTO keypoints VALUES (?, ?, ?, ?)\",\n            (image_id,) + keypoints.shape + (array_to_blob(keypoints),))\n\n    def add_descriptors(self, image_id, descriptors):\n        descriptors = np.ascontiguousarray(descriptors, np.uint8)\n        self.execute(\n            \"INSERT INTO descriptors VALUES (?, ?, ?, ?)\",\n            (image_id,) + descriptors.shape + (array_to_blob(descriptors),))\n\n    def add_matches(self, image_id1, image_id2, matches):\n        assert(len(matches.shape) == 2)\n        assert(matches.shape[1] == 2)\n\n        if image_id1 > image_id2:\n            matches = matches[:,::-1]\n\n        pair_id = image_ids_to_pair_id(image_id1, image_id2)\n        matches = np.asarray(matches, np.uint32)\n        self.execute(\n            \"INSERT INTO matches VALUES (?, ?, ?, ?)\",\n            (pair_id,) + matches.shape + (array_to_blob(matches),))\n\n    def add_two_view_geometry(self, image_id1, image_id2, matches,\n                            F=np.eye(3), E=np.eye(3), H=np.eye(3), config=2):\n        assert(len(matches.shape) == 2)\n        assert(matches.shape[1] == 2)\n\n        if image_id1 > image_id2:\n            matches = matches[:,::-1]\n\n        pair_id = image_ids_to_pair_id(image_id1, image_id2)\n        matches = np.asarray(matches, np.uint32)\n        F = np.asarray(F, dtype=np.float64)\n        E = np.asarray(E, dtype=np.float64)\n        H = np.asarray(H, dtype=np.float64)\n        self.execute(\n            \"INSERT INTO two_view_geometries VALUES (?, ?, ?, ?, ?, ?, ?, ?)\",\n            (pair_id,) + matches.shape + (array_to_blob(matches), config,\n            array_to_blob(F), array_to_blob(E), array_to_blob(H)))\n\nimport os, argparse, h5py, warnings\nimport numpy as np\nfrom tqdm import tqdm\nfrom PIL import Image, ExifTags\n\n\ndef get_focal(image_path, err_on_default=False):\n    image         = Image.open(image_path)\n    max_size      = max(image.size)\n\n    exif = image.getexif()\n    focal = None\n    if exif is not None:\n        focal_35mm = None\n        # https://github.com/colmap/colmap/blob/d3a29e203ab69e91eda938d6e56e1c7339d62a99/src/util/bitmap.cc#L299\n        for tag, value in exif.items():\n            focal_35mm = None\n            if ExifTags.TAGS.get(tag, None) == 'FocalLengthIn35mmFilm':\n                focal_35mm = float(value)\n                break\n\n        if focal_35mm is not None:\n            focal = focal_35mm / 35. * max_size\n    \n    if focal is None:\n        if err_on_default:\n            raise RuntimeError(\"Failed to find focal length\")\n\n        # failed to find it in exif, use prior\n        FOCAL_PRIOR = 1.2\n        focal = FOCAL_PRIOR * max_size\n\n    return focal\n\ndef create_camera(db, image_path, camera_model):\n    image         = Image.open(image_path)\n    width, height = image.size\n\n    focal = get_focal(image_path)\n\n    if camera_model == 'simple-pinhole':\n        model = 0 # simple pinhole\n        param_arr = np.array([focal, width / 2, height / 2])\n    if camera_model == 'pinhole':\n        model = 1 # pinhole\n        param_arr = np.array([focal, focal, width / 2, height / 2])\n    elif camera_model == 'simple-radial':\n        model = 2 # simple radial\n        param_arr = np.array([focal, width / 2, height / 2, 0.1])\n    elif camera_model == 'opencv':\n        model = 4 # opencv\n        param_arr = np.array([focal, focal, width / 2, height / 2, 0., 0., 0., 0.])\n        \n    return db.add_camera(model, width, height, param_arr)\n\n\ndef add_keypoints(db, h5_path, image_path, img_ext, camera_model, single_camera = True):\n    keypoint_f = h5py.File(os.path.join(h5_path, 'keypoints.h5'), 'r')\n\n    camera_id = None\n    fname_to_id = {}\n    for filename in tqdm(list(keypoint_f.keys())):\n        keypoints = keypoint_f[filename][()]\n\n        fname_with_ext = filename# + img_ext\n        path = os.path.join(image_path, fname_with_ext)\n        if not os.path.isfile(path):\n            raise IOError(f'Invalid image path {path}')\n\n        if camera_id is None or not single_camera:\n            camera_id = create_camera(db, path, camera_model)\n        image_id = db.add_image(fname_with_ext, camera_id)\n        fname_to_id[filename] = image_id\n\n        db.add_keypoints(image_id, keypoints)\n\n    return fname_to_id\n\ndef add_matches(db, h5_path, fname_to_id):\n    match_file = h5py.File(os.path.join(h5_path, 'matches.h5'), 'r')\n    \n    added = set()\n    n_keys = len(match_file.keys())\n    n_total = (n_keys * (n_keys - 1)) // 2\n\n    with tqdm(total=n_total) as pbar:\n        for key_1 in match_file.keys():\n            group = match_file[key_1]\n            for key_2 in group.keys():\n                id_1 = fname_to_id[key_1]\n                id_2 = fname_to_id[key_2]\n\n                pair_id = image_ids_to_pair_id(id_1, id_2)\n                if pair_id in added:\n                    warnings.warn(f'Pair {pair_id} ({id_1}, {id_2}) already added!')\n                    continue\n            \n                matches = group[key_2][()]\n                db.add_matches(id_1, id_2, matches)\n\n                added.add(pair_id)\n\n                pbar.update(1)\n                \ndef import_into_colmap(img_dir,\n                    feature_dir ='.featureout',\n                    database_path = 'colmap.db',\n                    img_ext='.jpg'):\n    db = COLMAPDatabase.connect(database_path)\n    db.create_tables()\n    single_camera = False\n    fname_to_id = add_keypoints(db, feature_dir, img_dir, img_ext, CONFIG.CAMERA_MODEL, single_camera)\n    add_matches(\n        db,\n        feature_dir,\n        fname_to_id,\n    )\n\n    db.commit()\n    return\n\nfrom torchvision.io import read_image as T_read_image\nfrom torchvision.io import ImageReadMode\nfrom torchvision import transforms as T\nfrom check_orientation.pre_trained_models import create_model\n\n# General utilities\nimport os\nfrom tqdm import tqdm\nfrom time import time\nfrom fastprogress import progress_bar\nimport gc\nimport numpy as np\nimport h5py\nfrom IPython.display import clear_output\nfrom collections import defaultdict\nfrom copy import deepcopy\nimport pandas as pd\nfrom dataclasses import dataclass\nimport gc\n# CV/ML\nimport cv2\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport kornia as K\nimport kornia.feature as KF\nfrom PIL import Image\nimport timm\nfrom timm.data import resolve_data_config\nfrom timm.data.transforms_factory import create_transform\nfrom torchvision import transforms\n\n# We will use ViT global descriptor to get matching shortlists.\ndef get_global_desc(fnames, model, model2,\n                    device =  device):\n    scalepic=600\n    model = model.eval()\n    model = model.to(device)\n    model2 = model2.eval()\n    model2 = model2.to(device)\n    config = resolve_data_config({}, model=model)\n    transform = transforms.Compose([\n                transforms.Resize(scalepic, interpolation=transforms.InterpolationMode.BICUBIC),\n                transforms.CenterCrop(scalepic),\n                transforms.ToTensor(),\n                transforms.Normalize([0.4850, 0.4560, 0.4060], [0.2290, 0.2240, 0.2250]),])#create_transform(**config)\n    global_descs_convnext=[]\n    for i, img_fname_full in tqdm(enumerate(fnames),total= len(fnames)):\n        key = os.path.splitext(os.path.basename(img_fname_full))[0]\n        img = Image.open(img_fname_full).convert('RGB')\n        timg = transform(img).unsqueeze(0).to(device)\n        with torch.no_grad():\n            desc = model.forward_features(timg.to(device)).mean(dim=(-1,2))\n            desc2 = model2.forward_features(timg.to(device)).mean(dim=(-1,2))\n            desc = desc.view(1, -1)\n            desc2 = desc2.view(1, -1)\n            desc_norm = torch.cat([desc, desc2], dim=-1)\n            desc_norm = F.normalize(desc_norm, dim=1, p=2)\n        global_descs_convnext.append(desc_norm.detach().cpu())\n    global_descs_all = torch.cat(global_descs_convnext, dim=0)\n    return global_descs_all\n\n\ndef get_img_pairs_exhaustive(img_fnames):\n    index_pairs = []\n    for i in range(len(img_fnames)):\n        for j in range(i+1, len(img_fnames)):\n            index_pairs.append((i,j))\n    return index_pairs\ndef get_image_pairs_shortlist(fnames,\n                            exhaustive_if_less = 20,\n                            nneighbor=60,\n                            device=torch.device('cpu'),\n                            th=0.1):\n    num_imgs = len(fnames)\n    if num_imgs <= exhaustive_if_less or num_imgs <= nneighbor :\n        return get_img_pairs_exhaustive(fnames)\n\n    model = timm.create_model('tf_efficientnet_b6',\n                            checkpoint_path='../input/tf-efficientnet/pytorch/tf-efficientnet-b6/1/tf_efficientnet_b6_aa-80ba17e4.pth')\n    model.eval()\n    model2 = timm.create_model('tf_efficientnet_b7',\n                            checkpoint_path='../input/tf-efficientnet/pytorch/tf-efficientnet-b7/1/tf_efficientnet_b7_ra-6c08e654.pth')                  \n    model2.eval()\n    descs = get_global_desc(fnames, model, model2, device=device)\n    del model, model2\n    torch.cuda.empty_cache()\n    import gc\n    gc.collect()\n    dm = torch.einsum('bi,ki->bk', descs, descs).detach().cpu() \n    value, index = torch.topk(dm, k=nneighbor, dim=1)\n\n    matching_list = []\n    for i in range(num_imgs-1):\n        for t in index[i][value[i]>th]:\n            if t == i:\n                continue\n            matching_list.append(tuple(sorted((i, t.item()))))\n    matching_list = sorted(list(set(matching_list)))\n    return matching_list\n\ndef load_torch_image(fname, device=torch.device('cpu')):\n    img = K.io.load_image(fname, K.io.ImageLoadType.RGB32, device=device)[None, ...]\n    return img\n\ndef convert_coord(r, w, h, rotk):\n    if rotk == 0:\n        return r\n    elif rotk == 1:\n        rx = w-1-r[:, 1]\n        ry = r[:, 0]\n        return torch.concat([rx[None], ry[None]], dim=0).T\n    elif rotk == 2:\n        rx = w-1-r[:, 0]\n        ry = h-1-r[:, 1]\n        return torch.concat([rx[None], ry[None]], dim=0).T\n    elif rotk == 3:\n        rx = r[:, 1]\n        ry = h-1-r[:, 0]\n        return torch.concat([rx[None], ry[None]], dim=0).T\n\ndef detect_common(img_fnames,\n                model_name,\n                rots,\n                file_keypoints,\n                feature_dir = '.featureout',\n                num_features = 4096,\n                resize_to = 1024,\n                detection_threshold = 0.005,\n                device=torch.device('cpu'),\n                min_matches=15,verbose=True\n                ):\n    if not os.path.isdir(feature_dir):\n        os.makedirs(feature_dir)\n    dict_model = {\n        \"aliked\" : ALIKED,\n        \"superpoint\" : SuperPoint,\n        \"doghardnet\" : DoGHardNet,\n        \"disk\" : DISK,\n        \"sift\" : SIFT,\n    }\n    extractor_class = dict_model[model_name]\n    \n    dtype = torch.float32 # ALIKED has issues with float16\n    extractor = extractor_class(\n        max_num_keypoints=num_features, detection_threshold=detection_threshold, resize=resize_to\n    ).eval().to(device, dtype)\n    rot=True\n    dict_kpts_cuda = {}\n    dict_descs_cuda = {}\n    for (img_path, rot_k) in zip(img_fnames, rots):\n        img_fname = img_path.split('/')[-1]\n        key = img_fname\n        with torch.inference_mode():\n            image0 = load_torch_image(img_path, device=device).to(dtype)\n            h, w = image0.shape[2], image0.shape[3]\n            image1 = torch.rot90(image0, rot_k, [2, 3])\n            feats0 = extractor.extract(image1)  # auto-resize the image, disable with resize=None\n            kpts = feats0['keypoints'].reshape(-1, 2).detach()\n            descs = feats0['descriptors'].reshape(len(kpts), -1).detach()\n            if rot==True:\n                kpts = convert_coord(kpts, w, h, rot_k)\n            dict_kpts_cuda[f\"{key}\"] = kpts\n            dict_descs_cuda[f\"{key}\"] = descs\n            print(f\"{model_name} > rot_k={rot_k}, kpts.shape={kpts.shape}, descs.shape={descs.shape}\")\n    del extractor\n    gc.collect()\n\n    #####################################################\n    # Matching keypoints\n    #####################################################\n    lg_matcher = KF.LightGlueMatcher(model_name, {\"width_confidence\": -1,\n                                            \"depth_confidence\": -1,\n                                            \"mp\": True if 'cuda' in str(device) else False}).eval().to(device)\n    \n    cnt_pairs = 0\n    with h5py.File(file_keypoints, mode='w') as f_match:\n        for pair_idx in tqdm(index_pairs):\n            idx1, idx2 = pair_idx\n            fname1, fname2 = img_fnames[idx1], img_fnames[idx2]\n            \n            key1, key2 = fname1.split('/')[-1], fname2.split('/')[-1]\n            \n            \n            kp1 = dict_kpts_cuda[key1]\n            kp2 = dict_kpts_cuda[key2]\n            \n            desc1 = dict_descs_cuda[key1]\n            desc2 = dict_descs_cuda[key2]\n            with torch.inference_mode():\n                dists, idxs = lg_matcher(desc1,\n                                    desc2,\n                                    KF.laf_from_center_scale_ori(kp1[None]),\n                                    KF.laf_from_center_scale_ori(kp2[None]))\n            if len(idxs)  == 0:\n                continue\n            n_matches = len(idxs)\n            kp1 = kp1[idxs[:,0], :].cpu().numpy().reshape(-1, 2).astype(np.float32)\n            kp2 = kp2[idxs[:,1], :].cpu().numpy().reshape(-1, 2).astype(np.float32)\n            group  = f_match.require_group(key1)\n            if n_matches >= min_matches:\n                group.create_dataset(key2, data=np.concatenate([kp1, kp2], axis=1))\n                cnt_pairs+=1\n                print (f'{model_name}> {key1}-{key2}: {n_matches} matches @ {cnt_pairs}th pair({model_name}+lightglue)')            \n            else:\n                print (f'{model_name}> {key1}-{key2}: {n_matches} matches --> skipped')\n    del lg_matcher\n    torch.cuda.empty_cache()\n    gc.collect()\n    return\n\ndef detect_lightglue_common(\n    img_fnames, model_name, index_pairs, feature_dir, device, file_keypoints, rots,\n    resize_to=1024,\n    detection_threshold=0.005, \n    num_features=4096, \n    min_matches=15,\n):\n    t=time()\n    detect_common(\n        img_fnames, model_name, rots, file_keypoints, feature_dir, \n        resize_to=resize_to,\n        num_features=num_features, \n        detection_threshold=detection_threshold, \n        device=device,\n        min_matches=min_matches,\n    )\n    gc.collect()\n    t=time() -t \n    print(f'Features matched in  {t:.4f} sec ({model_name}+LightGlue)')\n    return t\n\n\nimport sys\nsys.path.append(\"/kaggle/input/super-glue-pretrained-network\")\nfrom models.matching import Matching\nfrom models.superpoint import SuperPoint as SG_SuperPoint\nfrom models.superglue import SuperGlue\nfrom models.utils import (compute_pose_error, compute_epipolar_error,\n                        estimate_pose, make_matching_plot,\n                        error_colormap, AverageTimer, pose_auc, read_image,\n                        process_resize, frame2tensor,\n                        rotate_intrinsics, rotate_pose_inplane,\n                        scale_intrinsics)\n\nfrom torch.nn import functional as torchF  # For resizing tensor\n\n# Preprocess\ndef sg_read_image(image, device, resize):\n    w, h = image.shape[1], image.shape[0]\n    w_new, h_new = process_resize(w, h, [resize,])\n    \n    unit_shape = 8\n    w_new = w_new // unit_shape * unit_shape\n    h_new = h_new // unit_shape * unit_shape\n    \n    scales = (float(w) / float(w_new), float(h) / float(h_new))\n    image = cv2.resize(image.astype('float32'), (w_new, h_new))\n\n    inp = frame2tensor(image, \"cpu\")\n    return image, inp, scales, (h, w)\n\nclass SGDataset(Dataset):\n    def __init__(self, img_fnames, resize_to, device):\n        self.img_fnames = img_fnames\n        self.resize_to = resize_to\n        self.device = device\n        \n    def __len__(self):\n        return len(self.img_fnames)\n    \n    def __getitem__(self, idx):\n        fname = self.img_fnames[idx]\n        im = cv2.imread(fname, cv2.IMREAD_GRAYSCALE)\n        _, image, scale, ori_shape = sg_read_image(im, self.device, self.resize_to)\n        return image, torch.tensor([idx]), torch.tensor(ori_shape)\n\ndef get_superglue_dataloader(img_fnames, resize_to, device, batch_size=1):\n    dataset = SGDataset(img_fnames, resize_to, device)\n    dataloader = DataLoader(\n        dataset=dataset,\n        shuffle=False,\n        batch_size=batch_size,\n        pin_memory=True,\n        num_workers=2,\n        drop_last=False\n    )\n    return dataloader\n\ndef detect_superglue(\n    img_fnames, index_pairs, feature_dir, device, sg_config, file_keypoints, file_keypoints_crop,\n    resize_to=750, min_matches=15\n):    \n    t=time()\n\n    fnames1, fnames2, idxs1, idxs2 = [], [], [], []\n    for pair_idx in progress_bar(index_pairs):\n        idx1, idx2 = pair_idx\n        fname1, fname2 = img_fnames[idx1], img_fnames[idx2]\n        fnames1.append(fname1)\n        fnames2.append(fname2)\n        idxs1.append(idx1)\n        idxs2.append(idx2)\n        \n    dataloader = get_superglue_dataloader(img_fnames=img_fnames, resize_to=1680, device=device)\n\n    #####################################################\n    # Extract keypoints and descriptions\n    #####################################################\n    superpoint = SG_SuperPoint(sg_config[\"superpoint\"]).eval().to(device)\n    dict_features_cuda = {}\n    dict_shapes = {}\n    dict_images = {}\n    #dict_fname_shapes = {}\n    for X in dataloader:\n        image, idx, ori_shape = X\n        image = image[0].to(device)\n        fname = img_fnames[idx]\n        #dict_fname_shapes[fname] = ori_shape\n        key = fname.split('/')[-1]\n        \n        with torch.no_grad(), torch.autocast(device_type=\"cuda\", dtype=torch.float16):\n            pred = superpoint({'image': image})\n            dict_features_cuda[key] = pred\n            dict_shapes[key] = ori_shape\n            dict_images[key] = image.half()\n    del superpoint\n    gc.collect()\n    \n    #####################################################\n    # Matching keypoints\n    #####################################################\n    superglue = SuperGlue(sg_config[\"superglue\"]).eval().to(device)\n    weights = sg_config[\"superglue\"][\"weights\"]\n    cnt_pairs = 0\n    \n    mkpt = {}\n    with h5py.File(file_keypoints, mode='w') as f_match:\n        for idx, (fname1, fname2) in enumerate(zip(fnames1, fnames2)):\n            key1, key2 = fname1.split('/')[-1], fname2.split('/')[-1]\n\n            data = {\"image0\": dict_images[key1], \"image1\": dict_images[key2]}\n            data = {**data, **{k+'0': v for k, v in dict_features_cuda[key1].items()}}\n            data = {**data, **{k+'1': v for k, v in dict_features_cuda[key2].items()}}\n            for k in data:\n                if isinstance(data[k], (list, tuple)):\n                    data[k] = torch.stack(data[k])\n            with torch.no_grad(), torch.autocast(device_type=\"cuda\", dtype=torch.float16):\n                pred = {**data, **superglue(data)}\n                pred = {k: v[0].detach().cpu().numpy().copy() for k, v in pred.items()}\n            mkpts1, mkpts2 = pred[\"keypoints0\"], pred[\"keypoints1\"]\n            matches, conf = pred[\"matches0\"], pred[\"matching_scores0\"]\n            valid = matches > -1\n            mkpts1 = mkpts1[valid]\n            mkpts2 = mkpts2[matches[valid]]\n            mconf = conf[valid]\n            ori_shape_1 = dict_shapes[key1][0].numpy()\n            ori_shape_2 = dict_shapes[key2][0].numpy()\n            # Scaling coords\n            mkpts1[:,0] = mkpts1[:,0] * ori_shape_1[1] / dict_images[key1].shape[3]   # X\n            mkpts1[:,1] = mkpts1[:,1] * ori_shape_1[0] / dict_images[key1].shape[2]   # Y\n            mkpts2[:,0] = mkpts2[:,0] * ori_shape_2[1] / dict_images[key2].shape[3]   # X\n            mkpts2[:,1] = mkpts2[:,1] * ori_shape_2[0] / dict_images[key2].shape[2]   # Y  \n            n_matches = mconf.shape[0]\n            mkpt[fname1] = np.concatenate([mkpt[fname1], mkpts1], axis=0).astype(np.float32) if fname1 in mkpt else mkpts1\n            mkpt[fname2] = np.concatenate([mkpt[fname2], mkpts2], axis=0).astype(np.float32) if fname2 in mkpt else mkpts2\n            group  = f_match.require_group(key1)\n            if n_matches >= min_matches:\n                group.create_dataset(key2, data=np.concatenate([mkpts1, mkpts2], axis=1).astype(np.float32))\n                cnt_pairs+=1\n                print (f'{key1}-{key2}: {n_matches} matches @ {cnt_pairs}th pair(superglue/{resize_to}/{weights})')            \n            else:\n                print (f'{key1}-{key2}: {n_matches} matches --> skipped')\n    \n    \n    \n    gc.collect()\n    del superglue\n    del dict_features_cuda\n    del dict_images\n    torch.cuda.empty_cache()\n    gc.collect()\n    t=time() -t \n    print(f'Features matched in  {t:.4f} sec')\n    return t\n\n# Making kornia local features loading w/o internet\nclass AffNetHardNet(KF.LocalFeature):\n    \"\"\"Convenience module, which implements KeyNet detector + AffNet + HardNet descriptor.\n\n    .. image:: _static/img/keynet_affnet.jpg\n    \"\"\"\n\n    def __init__(\n        self,\n        num_features: int = 5000,\n        upright: bool = False,\n        device = torch.device('cpu'),\n        scale_laf: float = 1.0,\n        detector_type='GFTT',\n    ):\n        config = {\n            # Extraction Parameters\n            \"nms_size\": 15,\n            \"pyramid_levels\": 4,\n            \"up_levels\": 1,\n            \"scale_factor_levels\": math.sqrt(2),\n            \"s_mult\": 22.0,\n        }\n        ori_module = KF.PassLAF() if upright else KF.LAFOrienter(angle_detector=KF.OriNet(False)).eval()\n        if not upright:\n            weights = torch.load('/kaggle/input/kornia-local-feature-weights/OriNet.pth')['state_dict']\n            ori_module.angle_detector.load_state_dict(weights)\n        #detector = KF.KeyNetDetector(\n        #    False, num_features=num_features, ori_module=ori_module, aff_module=KF.LAFAffNetShapeEstimator(False).eval()\n        #).to(device)\n        if detector_type=='GFTT':\n            detector = KF.MultiResolutionDetector(\n                    KF.CornerGFTT(),\n                    num_features=num_features,\n                    config=config,\n                    ori_module=ori_module,\n                    aff_module=KF.LAFAffNetShapeEstimator(False).eval(),\n                ).to(device)\n        elif detector_type=='DoG':\n            detector = KF.MultiResolutionDetector(\n                    KF.BlobDoGSingle(),\n                    num_features=num_features,\n                    config=config,\n                    ori_module=ori_module,\n                    aff_module=KF.LAFAffNetShapeEstimator(False).eval(),\n                ).to(device)\n        else:\n            detector = KF.MultiResolutionDetector(\n                    KF.CornerHarris(0.04),\n                    num_features=num_features,\n                    config=config,\n                    ori_module=ori_module,\n                    aff_module=KF.LAFAffNetShapeEstimator(False).eval(),\n                ).to(device)\n        #kn_weights = torch.load('/kaggle/input/kornia-local-feature-weights/keynet_pytorch.pth')['state_dict']\n        #detector.model.load_state_dict(kn_weights)\n        affnet_weights = torch.load('/kaggle/input/kornia-local-feature-weights/AffNet.pth')['state_dict']\n        detector.aff.load_state_dict(affnet_weights)\n        \n        hardnet = KF.HardNet(False).eval()\n        hn_weights = torch.load('/kaggle/input/kornia-local-feature-weights/HardNetLib.pth')['state_dict']\n        hardnet.load_state_dict(hn_weights)\n        descriptor = KF.LAFDescriptor(hardnet, patch_size=32, grayscale_descriptor=True).to(device)\n        super().__init__(detector, descriptor, scale_laf)\n\n# Making kornia local features loading w/o internet\nclass KeyNetAffNetHardNet(KF.LocalFeature):\n    \"\"\"Convenience module, which implements KeyNet detector + AffNet + HardNet descriptor.\n\n    .. image:: _static/img/keynet_affnet.jpg\n    \"\"\"\n\n    def __init__(\n        self,\n        num_features: int = 5000,\n        upright: bool = False,\n        device = torch.device('cpu'),\n        scale_laf: float = 1.0,\n        \n    ):\n        ori_module = KF.PassLAF() if upright else KF.LAFOrienter(angle_detector=KF.OriNet(False)).eval()\n        if not upright:\n            weights = torch.load('/kaggle/input/kornia-local-feature-weights/OriNet.pth')['state_dict']\n            ori_module.angle_detector.load_state_dict(weights)\n        detector = KF.KeyNetDetector(\n            False, num_features=num_features, ori_module=ori_module, aff_module=KF.LAFAffNetShapeEstimator(False).eval()\n        ).to(device)\n        kn_weights = torch.load('/kaggle/input/kornia-local-feature-weights/keynet_pytorch.pth')['state_dict']\n        detector.model.load_state_dict(kn_weights)\n        affnet_weights = torch.load('/kaggle/input/kornia-local-feature-weights/AffNet.pth')['state_dict']\n        detector.aff.load_state_dict(affnet_weights)\n        \n        hardnet = KF.HardNet(False).eval()\n        hn_weights = torch.load('/kaggle/input/kornia-local-feature-weights/HardNetLib.pth')['state_dict']\n        hardnet.load_state_dict(hn_weights)\n        descriptor = KF.LAFDescriptor(hardnet, patch_size=32, grayscale_descriptor=True).to(device)\n        super().__init__(detector, descriptor, scale_laf)\n\n\ndef detect_features(img_fnames,\n                    num_feats = 8000,\n                    upright = False,\n                    device=torch.device('cpu'),\n                    feature_dir = '.featureout',\n                    resize_small_edge_to = 1200,\n                local_feature='Keynet'):\n\n    if local_feature == 'Keynet':\n        feature = KeyNetAffNetHardNet(num_feats, upright, device).to(device).eval()\n        torch.cuda.empty_cache()\n        gc.collect()\n    else:\n        feature = AffNetHardNet(num_feats, upright, device, detector_type=local_feature).to(device).eval()\n        torch.cuda.empty_cache()\n        gc.collect()\n    print('local feature:', local_feature)\n    if not os.path.isdir(feature_dir):\n        os.makedirs(feature_dir)\n    with h5py.File(f'{feature_dir}/lafs.h5', mode='w') as f_laf, \\\n        h5py.File(f'{feature_dir}/keypoints.h5', mode='w') as f_kp, \\\n        h5py.File(f'{feature_dir}/descriptors.h5', mode='w') as f_desc:\n        for img_path in progress_bar(img_fnames):\n            img_fname = img_path.split('/')[-1]\n            key = img_fname\n            with torch.inference_mode():\n                timg = load_torch_image(img_path, device=device)\n                H, W = timg.shape[2:]\n                if resize_small_edge_to is None:\n                    timg_resized = timg\n                else:\n                    timg_resized = K.geometry.resize(timg, resize_small_edge_to, antialias=True)\n                    torch.cuda.empty_cache()\n                    gc.collect()\n                    print(f'Resized {timg.shape} to {timg_resized.shape} (resize_small_edge_to={resize_small_edge_to})')\n                h, w = timg_resized.shape[2:]\n                \n                lafs, resps, descs = feature(K.color.rgb_to_grayscale(timg_resized))\n                lafs[:,:,0,:] *= float(W) / float(w)\n                lafs[:,:,1,:] *= float(H) / float(h)\n                desc_dim = descs.shape[-1]\n                kpts = KF.get_laf_center(lafs).reshape(-1, 2).detach().cpu().numpy()\n                descs = descs.reshape(-1, desc_dim).detach().cpu().numpy()\n                f_laf[key] = lafs.detach().cpu().numpy()\n                f_kp[key] = kpts\n                f_desc[key] = descs\n                del timg,timg_resized,lafs,kpts,descs\n                torch.cuda.empty_cache()\n                gc.collect()\n            \n            torch.cuda.empty_cache()\n            gc.collect()\n    del feature\n    torch.cuda.empty_cache()\n    gc.collect()\n    return\n\ndef match_features(img_fnames,\n                index_pairs,\n                file_keypoints,\n                feature_dir = '.featureout',\n                device=torch.device('cpu'),\n                min_matches=15, \n                force_mutual = True,\n                matching_alg='adalam'\n                ):\n    assert matching_alg in ['smnn', 'adalam']\n    with h5py.File(f'{feature_dir}/lafs.h5', mode='r') as f_laf, \\\n        h5py.File(f'{feature_dir}/keypoints.h5', mode='r') as f_kp, \\\n        h5py.File(f'{feature_dir}/descriptors.h5', mode='r') as f_desc, \\\n        h5py.File(file_keypoints, mode='w') as f_match:\n\n        for pair_idx in progress_bar(index_pairs):\n                    idx1, idx2 = pair_idx\n                    fname1, fname2 = img_fnames[idx1], img_fnames[idx2]\n                    key1, key2 = fname1.split('/')[-1], fname2.split('/')[-1]\n                    lafs1 = torch.from_numpy(f_laf[key1][...]).to(device)\n                    lafs2 = torch.from_numpy(f_laf[key2][...]).to(device)\n                    desc1 = torch.from_numpy(f_desc[key1][...]).to(device)\n                    desc2 = torch.from_numpy(f_desc[key2][...]).to(device)\n                    kp1 = torch.from_numpy(f_kp[key1][...]).to(device)\n                    kp2 = torch.from_numpy(f_kp[key2][...]).to(device)\n                    if matching_alg == 'adalam':\n                        img1, img2 = cv2.imread(fname1), cv2.imread(fname2)\n                        hw1, hw2 = img1.shape[:2], img2.shape[:2]\n                        adalam_config = KF.adalam.get_adalam_default_config()\n                        #adalam_config['orientation_difference_threshold'] = None\n                        #adalam_config['scale_rate_threshold'] = None\n                        adalam_config['force_seed_mnn']= True\n                        adalam_config['search_expansion'] = 32\n                        adalam_config['ransac_iters'] = 256\n                        adalam_config['device'] = device\n                        dists, idxs = KF.match_adalam(desc1, desc2,\n                                                    lafs1, lafs2, # Adalam takes into account also geometric information\n                                                    hw1=hw1, hw2=hw2,\n                                                    config=adalam_config) # Adalam also benefits from knowing image size\n                    else:\n                        dists, idxs = KF.match_smnn(desc1, desc2, 0.98)\n                    if len(idxs)  == 0:\n                        continue\n                    # Force mutual nearest neighbors\n                    if force_mutual:\n                        first_indices = get_unique_idxs(idxs[:,1])\n                        idxs = idxs[first_indices]\n                        dists = dists[first_indices]\n                    n_matches = len(idxs)\n                    kp1 = kp1[idxs[:,0], :].cpu().numpy().reshape(-1, 2).astype(np.float32)\n                    kp2 = kp2[idxs[:,1], :].cpu().numpy().reshape(-1, 2).astype(np.float32)\n                    if False:\n                        print (f'{key1}-{key2}: {n_matches} matches')\n                    group  = f_match.require_group(key1)\n                    if n_matches >= min_matches:\n                        group.create_dataset(key2, data=np.concatenate([kp1, kp2], axis=1))\n    torch.cuda.empty_cache()\n    gc.collect()\n    return\n\n\ndef get_unique_idxs(A, dim=0):\n    # https://stackoverflow.com/questions/72001505/how-to-get-unique-elements-and-their-firstly-appeared-indices-of-a-pytorch-tenso\n    unique, idx, counts = torch.unique(A, dim=dim, sorted=True, return_inverse=True, return_counts=True)\n    _, ind_sorted = torch.sort(idx, stable=True)\n    cum_sum = counts.cumsum(0)\n    cum_sum = torch.cat((torch.tensor([0],device=cum_sum.device), cum_sum[:-1]))\n    first_indices = ind_sorted[cum_sum]\n    return first_indices\n\ndef get_keypoint_from_h5(fp, key1, key2):\n    rc = -1\n    try:\n        kpts = np.array(fp[key1][key2])\n        rc = 0\n        return (rc, kpts)\n    except:\n        return (rc, None)\n\ndef get_keypoint_from_multi_h5(fps, key1, key2):\n    list_mkpts = []\n    for fp in fps:\n        rc, mkpts = get_keypoint_from_h5(fp, key1, key2)\n        if rc == 0:\n            list_mkpts.append(mkpts)\n    if len(list_mkpts) > 0:\n        list_mkpts = np.concatenate(list_mkpts, axis=0)\n    else:\n        list_mkpts = None\n    return list_mkpts\n\ndef matches_merger(\n    img_fnames,\n    index_pairs,\n    files_keypoints,\n    save_file,\n    feature_dir = 'featureout',\n    filter_FundamentalMatrix = False,\n    filter_iterations = 10,\n    filter_threshold = 8,\n):\n    # open h5 files\n    fps = [ h5py.File(file, mode=\"r\") for file in files_keypoints ]\n\n    with h5py.File(save_file, mode='w') as f_match:\n        counter = 0\n        for pair_idx in progress_bar(index_pairs):\n            idx1, idx2 = pair_idx\n            fname1, fname2 = img_fnames[idx1], img_fnames[idx2]\n            key1, key2 = fname1.split('/')[-1], fname2.split('/')[-1]\n\n            # extract keypoints\n            mkpts = get_keypoint_from_multi_h5(fps, key1, key2)\n            if mkpts is None:\n                print(f\"skipped key1={key1}, key2={key2}\")\n                continue\n\n            ori_size = mkpts.shape[0]\n            if mkpts.shape[0] < CONFIG.MERGE_PARAMS[\"min_matches\"]:\n                continue\n            \n            if filter_FundamentalMatrix:\n                store_inliers = { idx:0 for idx in range(mkpts.shape[0]) }\n                idxs = np.array(range(mkpts.shape[0]))\n                for iter in range(filter_iterations):\n                    try:\n                        Fm, inliers = cv2.findFundamentalMat(\n                            mkpts[:,:2], mkpts[:,2:4], cv2.USAC_MAGSAC, 0.15, 0.9999, 35000)\n                        if Fm is not None:\n                            inliers = inliers > 0\n                            inlier_idxs = idxs[inliers[:, 0]]\n                            #print(inliers.shape, inlier_idxs[:5])\n                            for idx in inlier_idxs:\n                                store_inliers[idx] += 1\n                    except:\n                        print(f\"Failed to cv2.findFundamentalMat. mkpts.shape={mkpts.shape}\")\n                inliers = np.array([ count for (idx, count) in store_inliers.items() ]) >= filter_threshold\n                mkpts = mkpts[inliers]\n                if mkpts.shape[0] < 15:\n                    print(f\"skipped key1={key1}, key2={key2}: mkpts.shape={mkpts.shape} after filtered.\")\n                    continue\n                #print(f\"filter_FundamentalMatrix: {len(store_inliers)} matches --> {mkpts.shape[0]} matches\")\n            \n            \n            print (f'{key1}-{key2}: {ori_size} --> {mkpts.shape[0]} matches')            \n            # regist tmp file\n            group  = f_match.require_group(key1)\n            group.create_dataset(key2, data=mkpts)\n            counter += 1\n    print( f\"Ensembled pairs : {counter} pairs\" )\n    for fp in fps:\n        fp.close()\n\ndef keypoints_merger(\n    img_fnames,\n    index_pairs,\n    files_keypoints,\n    feature_dir = 'featureout',\n    filter_FundamentalMatrix = False,\n    filter_iterations = 10,\n    filter_threshold = 8,\n):\n    save_file = f'{feature_dir}/merge_tmp.h5'\n    !rm -rf {save_file}\n    matches_merger(\n        img_fnames,\n        index_pairs,\n        files_keypoints,\n        save_file,\n        feature_dir = feature_dir,\n        filter_FundamentalMatrix = filter_FundamentalMatrix,\n        filter_iterations = filter_iterations,\n        filter_threshold = filter_threshold,\n    )\n        \n    # Let's find unique loftr pixels and group them together.\n    kpts = defaultdict(list)\n    match_indexes = defaultdict(dict)\n    total_kpts=defaultdict(int)\n    with h5py.File(save_file, mode='r') as f_match:\n        for k1 in f_match.keys():\n            group  = f_match[k1]\n            for k2 in group.keys():\n                matches = group[k2][...]\n                total_kpts[k1]\n                kpts[k1].append(matches[:, :2])\n                kpts[k2].append(matches[:, 2:])\n                current_match = torch.arange(len(matches)).reshape(-1, 1).repeat(1, 2)\n                current_match[:, 0]+=total_kpts[k1]\n                current_match[:, 1]+=total_kpts[k2]\n                total_kpts[k1]+=len(matches)\n                total_kpts[k2]+=len(matches)\n                match_indexes[k1][k2]=current_match\n\n    for k in kpts.keys():\n        kpts[k] = np.round(np.concatenate(kpts[k], axis=0))\n    unique_kpts = {}\n    unique_match_idxs = {}\n    out_match = defaultdict(dict)\n    for k in kpts.keys():\n        uniq_kps, uniq_reverse_idxs = torch.unique(torch.from_numpy(kpts[k]),dim=0, return_inverse=True)\n        unique_match_idxs[k] = uniq_reverse_idxs\n        unique_kpts[k] = uniq_kps.numpy()\n    for k1, group in match_indexes.items():\n        for k2, m in group.items():\n            m2 = deepcopy(m)\n            m2[:,0] = unique_match_idxs[k1][m2[:,0]]\n            m2[:,1] = unique_match_idxs[k2][m2[:,1]]\n            mkpts = np.concatenate([unique_kpts[k1][ m2[:,0]],\n                                    unique_kpts[k2][  m2[:,1]],\n                                ],\n                                axis=1)\n            unique_idxs_current = get_unique_idxs(torch.from_numpy(mkpts), dim=0)\n            m2_semiclean = m2[unique_idxs_current]\n            unique_idxs_current1 = get_unique_idxs(m2_semiclean[:, 0], dim=0)\n            m2_semiclean = m2_semiclean[unique_idxs_current1]\n            unique_idxs_current2 = get_unique_idxs(m2_semiclean[:, 1], dim=0)\n            m2_semiclean2 = m2_semiclean[unique_idxs_current2]\n            out_match[k1][k2] = m2_semiclean2.numpy()\n    with h5py.File(f'{feature_dir}/keypoints.h5', mode='w') as f_kp:\n        for k, kpts1 in unique_kpts.items():\n            f_kp[k] = kpts1\n    \n    with h5py.File(f'{feature_dir}/matches.h5', mode='w') as f_match:\n        for k1, gr in out_match.items():\n            group  = f_match.require_group(k1)\n            for k2, match in gr.items():\n                group[k2] = match\n    return\n\ndef wrapper_keypoints(\n    img_fnames, index_pairs, feature_dir, device, timings, rots\n):\n    #############################################################\n    # get keypoints\n    #############################################################\n    files_keypoints = []\n    \n    local_feature_model = 'DoG' #Keynet GFTT DoG Harris\n    detect_features(img_fnames, \n                    16000,\n                    feature_dir=local_feature_model,\n                    upright=False,\n                    device=device,\n                    resize_small_edge_to=1600,\n                    local_feature=local_feature_model,#GFTT #DoG #Harris\n                            )\n    torch.cuda.empty_cache()\n    gc.collect()\n    file_keypoints = f\"{feature_dir}/matches_{local_feature_model}.h5\"\n    match_features(img_fnames, index_pairs, file_keypoints, feature_dir=local_feature_model,device=device)\n    files_keypoints.append( file_keypoints )\n    torch.cuda.empty_cache()\n    gc.collect()\n    \n#     local_feature_model = 'Keynet' #Keynet GFTT DoG Harris\n#     detect_features(img_fnames, \n#                     19000,\n#                     feature_dir=local_feature_model,\n#                     upright=False,\n#                     device=device,\n#                     resize_small_edge_to=1200,\n#                     local_feature=local_feature_model,#GFTT #DoG #Harris\n#                             )\n#     torch.cuda.empty_cache()\n#     gc.collect()\n#     file_keypoints = f\"{feature_dir}/matches_{local_feature_model}.h5\"\n#     match_features(img_fnames, index_pairs, file_keypoints, feature_dir=local_feature_model,device=device)\n#     files_keypoints.append( file_keypoints )\n#     torch.cuda.empty_cache()\n#     gc.collect()\n\n    \n    if CONFIG.use_superglue:\n        for params_sg in CONFIG.params_sgs:\n            resize_to = params_sg[\"resize_to\"]\n            file_keypoints = f\"{feature_dir}/matches_superglue_{resize_to}pix.h5\"\n            file_keypoints_crop = f\"{feature_dir}/matches_superglue_{resize_to}pix_crop.h5\"\n            !rm -rf {file_keypoints}\n            t = detect_superglue(\n                img_fnames, index_pairs, feature_dir, device, \n                params_sg[\"sg_config\"], file_keypoints, file_keypoints_crop,\n                resize_to=params_sg[\"resize_to\"], \n                min_matches=params_sg[\"min_matches\"],\n            )\n            gc.collect()\n            files_keypoints.append( file_keypoints )\n            #files_keypoints.append( file_keypoints_crop )\n            timings['feature_matching'].append(t)\n\n    if CONFIG.use_aliked_lightglue:\n        model_name = \"aliked\"\n        file_keypoints = f'{feature_dir}/matches_lightglue_{model_name}.h5'\n        t = detect_lightglue_common(\n            img_fnames, model_name, index_pairs, feature_dir, device, file_keypoints, rots,\n            resize_to=CONFIG.params_aliked_lightglue[\"resize_to\"],\n            detection_threshold=CONFIG.params_aliked_lightglue[\"detection_threshold\"],\n            num_features=CONFIG.params_aliked_lightglue[\"num_features\"],\n            min_matches=CONFIG.params_aliked_lightglue[\"min_matches\"],\n        )\n        gc.collect()\n        files_keypoints.append(file_keypoints)\n        timings['feature_matching'].append(t)\n\n    if CONFIG.use_doghardnet_lightglue:\n        model_name = \"doghardnet\"\n        file_keypoints = f'{feature_dir}/matches_lightglue_{model_name}.h5'\n        t = detect_lightglue_common(\n            img_fnames, model_name, index_pairs, feature_dir, device, file_keypoints, rots,\n            resize_to=CONFIG.params_doghardnet_lightglue[\"resize_to\"],\n            detection_threshold=CONFIG.params_doghardnet_lightglue[\"detection_threshold\"],\n            num_features=CONFIG.params_doghardnet_lightglue[\"num_features\"],\n            min_matches=CONFIG.params_doghardnet_lightglue[\"min_matches\"],\n        )\n        gc.collect()\n        files_keypoints.append(file_keypoints)\n        timings['feature_matching'].append(t)\n\n    if CONFIG.use_superpoint_lightglue:\n        model_name = \"superpoint\"\n        file_keypoints = f'{feature_dir}/matches_lightglue_{model_name}.h5'\n        \n        t = detect_lightglue_common(\n            img_fnames, model_name, index_pairs, feature_dir, device, file_keypoints, rots,\n            resize_to=CONFIG.params_superpoint_lightglue[\"resize_to\"],\n            detection_threshold=CONFIG.params_superpoint_lightglue[\"detection_threshold\"],\n            num_features=CONFIG.params_superpoint_lightglue[\"num_features\"],\n            min_matches=CONFIG.params_superpoint_lightglue[\"min_matches\"],\n        )\n        gc.collect()\n        files_keypoints.append(file_keypoints)\n        \n        timings['feature_matching'].append(t)\n\n\n    #############################################################\n    # merge keypoints\n    #############################################################\n    keypoints_merger(\n        img_fnames,\n        index_pairs,\n        files_keypoints,\n        feature_dir = feature_dir,\n        filter_FundamentalMatrix = CONFIG.MERGE_PARAMS[\"filter_FundamentalMatrix\"],\n        filter_iterations = CONFIG.MERGE_PARAMS[\"filter_iterations\"],\n        filter_threshold = CONFIG.MERGE_PARAMS[\"filter_threshold\"],\n    )    \n    return timings\n\ndef reconstruct_from_db(dataset, scene, feature_dir, img_dir, timings, image_paths):\n    scene_result = {}\n    #############################################################\n    # regist keypoints from h5 into colmap db\n    #############################################################\n    database_path = f'{feature_dir}/colmap.db'\n    if os.path.isfile(database_path):\n        os.remove(database_path)\n    gc.collect()\n    import_into_colmap(img_dir, feature_dir=feature_dir, database_path=database_path)\n    output_path = f'{feature_dir}/colmap_rec'\n\n    #############################################################\n    # Calculate fundamental matrix with colmap api\n    #############################################################\n    t=time()\n    options = pycolmap.SiftMatchingOptions()\n    #options.confidence = 0.9999\n    #options.max_num_trials = 20000\n    pycolmap.match_exhaustive(database_path, sift_options=options)\n    t=time() - t \n    timings['RANSAC'].append(t)\n    print(f'RANSAC in  {t:.4f} sec')\n\n    #############################################################\n    # Execute bundle adjustmnet with colmap api\n    # --> Bundle adjustment Calcs Camera matrix, R and t\n    #############################################################\n    t=time()\n    # By default colmap does not generate a reconstruction if less than 10 images are registered. Lower it to 3.\n    mapper_options = pycolmap.IncrementalMapperOptions()\n    #mapper_options.num_threads = 1\n    mapper_options.min_model_size = 3\n    os.makedirs(output_path, exist_ok=True)\n    maps = pycolmap.incremental_mapping(database_path=database_path, image_path=img_dir, output_path=output_path, options=mapper_options)\n    torch.cuda.empty_cache()\n    gc.collect()\n    print(maps)\n    clear_output(wait=False)\n    t=time() - t\n    timings['Reconstruction'].append(t)\n    print(f'Reconstruction done in  {t:.4f} sec')\n\n    #############################################################\n    # Extract R,t from maps \n    #############################################################            \n    imgs_registered  = 0\n    best_idx = None\n    list_num_images = []            \n    print (\"Looking for the best reconstruction\")\n    if isinstance(maps, dict):\n        for idx1, rec in maps.items():\n            print (idx1, rec.summary())\n            list_num_images.append( len(rec.images) )\n            if len(rec.images) > imgs_registered:\n                imgs_registered = len(rec.images)\n                best_idx = idx1\n    list_num_images = np.array(list_num_images)\n    print(f\"list_num_images = {list_num_images}\")\n    if best_idx is not None:\n        print (maps[best_idx].summary())\n        for k, im in maps[best_idx].images.items():\n            key1 = f'test/{dataset}/images/{im.name}'\n            scene_result[key1] = {}\n            scene_result[key1][\"R\"] = deepcopy(im.rotmat())\n            scene_result[key1][\"t\"] = deepcopy(np.array(im.tvec))\n            torch.cuda.empty_cache()\n            gc.collect()\n\n    print(f'Registered: {dataset} / {scene} -> {len(scene_result)} images')\n    print(f'Total: {dataset} / {scene} -> {len(image_paths)} images')\n    print(timings)\n    torch.cuda.empty_cache()\n    gc.collect()\n    return scene_result\n\ndef arr_to_str(a):\n    return ';'.join([str(x) for x in a.reshape(-1)])\n\n# Function to create a submission file.\ndef create_submission(out_results, data_dict):\n    with open(f'submission.csv', 'w') as f:\n        f.write('image_path,dataset,scene,rotation_matrix,translation_vector\\n')\n        for dataset in data_dict:\n            if dataset in out_results:\n                res = out_results[dataset]\n            else:\n                res = {}\n            for scene in data_dict[dataset]:\n                if scene in res:\n                    scene_res = res[scene]\n                else:\n                    scene_res = {\"R\":{}, \"t\":{}}\n                for image in data_dict[dataset][scene]:\n                    if image in scene_res:\n                        print (image)\n                        R = scene_res[image]['R'].reshape(-1)\n                        T = scene_res[image]['t'].reshape(-1)\n                    else:\n                        R = np.eye(3).reshape(-1)\n                        T = np.zeros((3))\n                    f.write(f'{image},{dataset},{scene},{arr_to_str(R)},{arr_to_str(T)}\\n')\n\nsrc = '/kaggle/input/image-matching-challenge-2024'\n\n# Get data from csv.\ndata_dict = {}\nwith open(f'{src}/sample_submission.csv', 'r') as f:\n    for i, l in enumerate(f):\n        # Skip header.\n        if l and i > 0:\n            image, dataset, scene, _, _ = l.strip().split(',')\n            if dataset not in data_dict:\n                data_dict[dataset] = {}\n            if scene not in data_dict[dataset]:\n                data_dict[dataset][scene] = []\n            data_dict[dataset][scene].append(image)\n            \n            if CONFIG.DRY_RUN:\n                if len(data_dict[dataset][scene]) == CONFIG.DRY_RUN_MAX_IMAGES:\n                    break\n                    \nfor dataset in data_dict:\n    for scene in data_dict[dataset]:\n        print(f'{dataset} / {scene} -> {len(data_dict[dataset][scene])} images')\n\nout_results = {}\ntimings = {\n    \"rotation_detection\" : [],\n    \"shortlisting\":[],\n\"feature_detection\": [],\n\"feature_matching\":[],\n\"RANSAC\": [],\n\"Reconstruction\": []\n}\n\ngc.collect()\ndatasets = []\nfor dataset in data_dict:\n    datasets.append(dataset)\n\nwith concurrent.futures.ProcessPoolExecutor(max_workers=CONFIG.NUM_CORES) as executors:\n    futures = defaultdict(dict)\n    for dataset in datasets:\n        print(dataset)\n        if dataset not in out_results:\n            out_results[dataset] = {}\n        for scene in data_dict[dataset]:\n            print(scene)\n            # Fail gently if the notebook has not been submitted and the test data is not populated.\n            # You may want to run this on the training data in that case?\n            img_dir = f'{src}/test/{dataset}/images'\n            if not os.path.exists(img_dir):\n                continue\n\n            out_results[dataset][scene] = {}\n            img_fnames = [f'{src}/{x}' for x in data_dict[dataset][scene]]\n            print (f\"Got {len(img_fnames)} images\")\n            feature_dir = f'featureout/{dataset}_{scene}'\n            if not os.path.isdir(feature_dir):\n                os.makedirs(feature_dir, exist_ok=True)\n\n            #############################################################\n            # get image rotations\n            #############################################################\n            t = time()\n            rots = [ 0 for fname in img_fnames ]\n            t = time()-t\n            timings['rotation_detection'].append(t)\n            print (f'rotation_detection for {len(img_fnames)} images : {t:.4f} sec')\n            gc.collect()\n            \n            #############################################################\n            # get image pairs\n            #############################################################\n            t=time()\n\n            index_pairs = get_image_pairs_shortlist(img_fnames,\n                                                nneighbor=40,              \n                                                exhaustive_if_less = 40,th=0.3)\n            t=time() -t \n            timings['shortlisting'].append(t)\n            print (f'{len(index_pairs)}, pairs to match, {t:.4f} sec')\n            torch.cuda.empty_cache()\n            gc.collect()\n\n            #############################################################\n            # get keypoints\n            #############################################################            \n            keypoints_timings = wrapper_keypoints(\n                img_fnames, index_pairs, feature_dir, device, timings, rots\n            )\n            torch.cuda.empty_cache()\n            gc.collect()\n            timings['feature_matching'] = keypoints_timings['feature_matching']\n\n            #############################################################\n            # kick COLMAP reconstruction\n            #############################################################            \n            futures[dataset][scene] = executors.submit(\n                reconstruct_from_db, \n                dataset, scene, feature_dir, img_dir, timings, data_dict[dataset][scene])\n                \n    #############################################################\n    # reconstruction results\n    #############################################################            \n    for dataset in datasets:\n        for scene in data_dict[dataset]:\n            # wait to complete COLMAP reconstruction\n            result = futures[dataset][scene].result()\n            if result is not None:\n                out_results[dataset][scene] = result   # get R and t from result\n            torch.cuda.empty_cache()\n            gc.collect()\n    create_submission(out_results, data_dict)\n    torch.cuda.empty_cache()\n    gc.collect()\n ","metadata":{"_kg_hide-input":true,"_kg_hide-output":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!cat submission.csv","metadata":{"_kg_hide-input":true,"_kg_hide-output":true},"execution_count":null,"outputs":[]}]}