![\"](\"https://i.imgur.com/S5xBDPg.png\")
\n
![\"](\"https://i.imgur.com/VBLFUdO.png\")
Simple Copy-Paste is a Strong Data Augmentation Method\nfor Instance Segmentation
\n
\n
Edited::
\nSimpleCopy Paste Augmentation Implementation in MMDetection
\nhttps://github.com/open-mmlab/mmdetection/pull/6282
\n\ntrain_pipeline = [\n dict(type='SimpleCopyPaste', prob=0),\n dict(\n type='Resize',\n img_scale=[(1333, 640), (1333, 672), (1333, 704), (1333, 736),\n (1333, 768), (1333, 800)],\n multiscale_mode='value',\n keep_ratio=True),\n dict(type='RandomCrop', crop_size=(1024, 1024)),\n dict(type='RandomFlip', flip_ratio=0.5),\n dict(type='Normalize', **img_norm_cfg),\n dict(type='Pad', size_divisor=32),\n dict(type='DefaultFormatBundle'),\n dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks'])\n]
To compose transforms with copy-paste augmentation with Albu::
\nhttps://github.com/conradry/copy-paste-aug
\n`import albumentations as A
\nfrom albumentations.pytorch.transforms import ToTensorV2
\nfrom copy_paste import CopyPaste
transform = A.Compose([
\n A.RandomScale(scale_limit=(-0.9, 1), p=1), #LargeScaleJitter from scale of 0.1 to 2
\n A.PadIfNeeded(256, 256, border_mode=0), #constant 0 border
\n A.RandomCrop(256, 256),
\n A.HorizontalFlip(p=0.5),
\n CopyPaste(blend=True, sigma=1, pct_objects_paste=0.5, p=1)
\n ], bbox_params=A.BboxParams(format=\"coco\")
\n)`
Simple Copy-Paste is a Strong Data Augmentation Method\nfor Instance Segmentation
\n
\n
Edited::
\nSimpleCopy Paste Augmentation Implementation in MMDetection
\nhttps://github.com/open-mmlab/mmdetection/pull/6282
\n\ntrain_pipeline = [\n dict(type='SimpleCopyPaste', prob=0),\n dict(\n type='Resize',\n img_scale=[(1333, 640), (1333, 672), (1333, 704), (1333, 736),\n (1333, 768), (1333, 800)],\n multiscale_mode='value',\n keep_ratio=True),\n dict(type='RandomCrop', crop_size=(1024, 1024)),\n dict(type='RandomFlip', flip_ratio=0.5),\n dict(type='Normalize', **img_norm_cfg),\n dict(type='Pad', size_divisor=32),\n dict(type='DefaultFormatBundle'),\n dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks'])\n]
To compose transforms with copy-paste augmentation with Albu::
\nhttps://github.com/conradry/copy-paste-aug
\n`import albumentations as A
\nfrom albumentations.pytorch.transforms import ToTensorV2
\nfrom copy_paste import CopyPaste
transform = A.Compose([
\n A.RandomScale(scale_limit=(-0.9, 1), p=1), #LargeScaleJitter from scale of 0.1 to 2
\n A.PadIfNeeded(256, 256, border_mode=0), #constant 0 border
\n A.RandomCrop(256, 256),
\n A.HorizontalFlip(p=0.5),
\n CopyPaste(blend=True, sigma=1, pct_objects_paste=0.5, p=1)
\n ], bbox_params=A.BboxParams(format=\"coco\")
\n)`
I actually tried copy-paste augmentation (I'm using Detectron2, it is quite easy to implement). I see a slight improvement, but not that much.
", "rawMarkdown": "I actually tried copy-paste augmentation (I'm using Detectron2, it is quite easy to implement). I see a slight improvement, but not that much.", "votes": 4, "replies": [ { "id": 1601138, "postDate": "2021-12-01T03:54:23.010Z", "content": "How do you integrate Simple Copy Paste with detectron2 ?
", "rawMarkdown": "How do you integrate Simple Copy Paste with detectron2 ?", "votes": 1 }, { "id": 1601162, "postDate": "2021-12-01T04:30:43.357Z", "content": "@namgalielei I create a custom mapper as in this documentation, and call copy-paste augmentation from that mapper: https://detectron2.readthedocs.io/en/latest/tutorials/data_loading.html
\nHere is the full code of my copy-paste augmentator using detectron2 datasets:
\nhttps://gist.github.com/hav4ik/2d80b63dcab69651ad7bf1495053e35c
The constructor takes the following parameters:
\nd2_dataset - basically a detectron2 dataset (returned by DatasetCatalog.get(name). I set it to my sartorius train set (without LiveCell samples)paste_same_class flag if set to True will only paste from a random sample of the same class.paste_density dictates what portion of cells in that random sample we chose will be copied over to the sample that we are augmenting. I usually set it to no more than 0.4. You can set a range.filter_area_thresh basically says \"if after copy-pasting some cells will have area less than that amount of their original size, we remove the annotations for that cell\".p is the probability of applying copy-paste augmentation.The __call__ method is quite straightforward. It just takes a dataset sample (again, in Detectron2 format) and applies augmentation to it. Notice I don't apply augmentations to samples in LiveCell collection.
@chankhavu thanks for your works. You have a good engineering skill.
", "rawMarkdown": "@chankhavu thanks for your works. You have a good engineering skill.", "votes": 3 }, { "id": 1601761, "postDate": "2021-12-01T15:06:23.210Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1601919, "postDate": "2021-12-01T16:38:05.810Z", "content": "self.cls_indices = [
\n[
\nI for I, item in enumerate(d2_dataset)
\nIf the item [' annotations'] [0] [' category_id] = = cls_index
\n]
\nFor cls_index in range (3)
\n]
\nWhy the error occurs:
\nTypeError: string indices must be integers, not str
for i, item in enumerate(d2_dataset)
\nI found that the contents of the loop were just keys in the dict. Is there a problem with the code?
@joker d2_dataset variable is supposed to be a list of dicts created using register_coco_instances as described in the documentation: https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html
This notebook by @slawekbiel is the ideal example of how to read the Sartorius dataset in that format (in that notebook, it is train_ds variable): https://www.kaggle.com/slawekbiel/positive-score-with-detectron-2-3-training
I want to learn how your Mapper is defined?
", "rawMarkdown": "I want to learn how your Mapper is defined?" }, { "id": 1603986, "postDate": "2021-12-03T02:06:37.023Z", "content": "As a newbie, I'd also like to know how your Mapper is defined and able to run successfully.
", "rawMarkdown": "As a newbie, I'd also like to know how your Mapper is defined and able to run successfully." }, { "id": 1603987, "postDate": "2021-12-03T02:12:13.630Z", "content": "Hello, would you share your defined Mapper, looking for your reply
", "rawMarkdown": "Hello, would you share your defined Mapper, looking for your reply" }, { "id": 1603996, "postDate": "2021-12-03T02:30:52.027Z", "content": "@zaopolearning hey, I added the mapper in this gist: https://gist.github.com/hav4ik/2d80b63dcab69651ad7bf1495053e35c - basically almost the same as the mapper in detectron documentation: https://detectron2.readthedocs.io/en/latest/tutorials/data_loading.html
\n(as you can see I use detectron2's transforms for geometric transforms like resizing and cropping, and albumentation for enhancing the image but…… augmentations doesn't seem to work here)
\nsorry for messy code lol :)
", "rawMarkdown": "@zaopolearning hey, I added the mapper in this gist: https://gist.github.com/hav4ik/2d80b63dcab69651ad7bf1495053e35c - basically almost the same as the mapper in detectron documentation: https://detectron2.readthedocs.io/en/latest/tutorials/data_loading.html\n\n(as you can see I use detectron2's transforms for geometric transforms like resizing and cropping, and albumentation for enhancing the image but...... augmentations doesn't seem to work here)\n\nsorry for messy code lol :) " }, { "id": 1604006, "postDate": "2021-12-03T02:57:01.307Z", "content": "thank you!
", "rawMarkdown": "thank you!" }, { "id": 1604012, "postDate": "2021-12-03T03:12:17.633Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1606219, "postDate": "2021-12-04T22:37:35.390Z", "content": "Hey @chankhavu ,thankyou for sharing this
\nI was trying to implement this and am getting the error in this part of the code:
\norig_img = cv2.imread(dataset_dict[\"file_name\"])
TypeError: list indices must be integers or slices, not str
\nMy dataset_dict is also a list of dicts (as you gave example of train_ds).
\nI don't know how to resolve this.
It doesn't seem to work when my images are different sizes.
", "rawMarkdown": "It doesn't seem to work when my images are different sizes." } ] }, { "id": 1607980, "postDate": "2021-12-06T07:01:15.433Z", "content": "does anyone tried this method in mmdetection?
", "rawMarkdown": "does anyone tried this method in mmdetection?", "votes": 1 }, { "id": 1601005, "postDate": "2021-11-30T23:18:55.440Z", "content": "Thanks @mlneo07. Were you able to try this out? Looks like mmdet hasnt merged this feature yet
", "rawMarkdown": "Thanks @mlneo07. Were you able to try this out? Looks like mmdet hasnt merged this feature yet", "votes": 1, "replies": [ { "id": 1601346, "postDate": "2021-12-01T07:23:05.787Z", "content": "Yes, Slight improvement. I am using albumentations. MMDet is not merged yet but we can try @PIPELINES.register_module() for cusotmize dataset.
", "rawMarkdown": "Yes, Slight improvement. I am using albumentations. MMDet is not merged yet but we can try [@PIPELINES.register_module()](https://mmdetection.readthedocs.io/en/latest/tutorials/data_pipeline.html) for cusotmize dataset. ", "votes": 1 }, { "id": 1601374, "postDate": "2021-12-01T08:12:38.030Z", "content": "thanks @mlneo07 . Do you have an example of the custom pipeline file? I tried adding the decorator to the CopyPaste class but I think that is not the right approach.
Do you use copy-paste in MMDetection?
", "rawMarkdown": "Do you use copy-paste in MMDetection?" } ] } ], "comments": [ { "id": 1601125, "author_name": "Chan Kha Vu", "author_url": "", "post_date": "2021-12-01T03:23:05.697000", "content": "I actually tried copy-paste augmentation (I'm using Detectron2, it is quite easy to implement). I see a slight improvement, but not that much.
", "votes": 4, "replies": [ { "id": 1601138, "author_name": "Liam Nguyen", "author_url": "", "post_date": "2021-12-01T03:54:23.010000", "content": "How do you integrate Simple Copy Paste with detectron2 ?
", "votes": 1, "replies": [] }, { "id": 1601162, "author_name": "Chan Kha Vu", "author_url": "", "post_date": "2021-12-01T04:30:43.357000", "content": "@namgalielei I create a custom mapper as in this documentation, and call copy-paste augmentation from that mapper: https://detectron2.readthedocs.io/en/latest/tutorials/data_loading.html
\nHere is the full code of my copy-paste augmentator using detectron2 datasets:
\nhttps://gist.github.com/hav4ik/2d80b63dcab69651ad7bf1495053e35c
The constructor takes the following parameters:
\nd2_dataset - basically a detectron2 dataset (returned by DatasetCatalog.get(name). I set it to my sartorius train set (without LiveCell samples)paste_same_class flag if set to True will only paste from a random sample of the same class.paste_density dictates what portion of cells in that random sample we chose will be copied over to the sample that we are augmenting. I usually set it to no more than 0.4. You can set a range.filter_area_thresh basically says \"if after copy-pasting some cells will have area less than that amount of their original size, we remove the annotations for that cell\".p is the probability of applying copy-paste augmentation.The __call__ method is quite straightforward. It just takes a dataset sample (again, in Detectron2 format) and applies augmentation to it. Notice I don't apply augmentations to samples in LiveCell collection.
@chankhavu thanks for your works. You have a good engineering skill.
", "votes": 3, "replies": [] }, { "id": 1601761, "author_name": "", "author_url": "", "post_date": "2021-12-01T15:06:23.210000", "content": "", "votes": 0, "replies": [] }, { "id": 1601919, "author_name": "yerongg", "author_url": "", "post_date": "2021-12-01T16:38:05.810000", "content": "self.cls_indices = [
\n[
\nI for I, item in enumerate(d2_dataset)
\nIf the item [' annotations'] [0] [' category_id] = = cls_index
\n]
\nFor cls_index in range (3)
\n]
\nWhy the error occurs:
\nTypeError: string indices must be integers, not str
for i, item in enumerate(d2_dataset)
\nI found that the contents of the loop were just keys in the dict. Is there a problem with the code?
@joker d2_dataset variable is supposed to be a list of dicts created using register_coco_instances as described in the documentation: https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html
This notebook by @slawekbiel is the ideal example of how to read the Sartorius dataset in that format (in that notebook, it is train_ds variable): https://www.kaggle.com/slawekbiel/positive-score-with-detectron-2-3-training
I want to learn how your Mapper is defined?
", "votes": 0, "replies": [] }, { "id": 1603986, "author_name": "iwrbeverly", "author_url": "", "post_date": "2021-12-03T02:06:37.023000", "content": "As a newbie, I'd also like to know how your Mapper is defined and able to run successfully.
", "votes": 0, "replies": [] }, { "id": 1603987, "author_name": "sanmedbio.huajia", "author_url": "", "post_date": "2021-12-03T02:12:13.630000", "content": "Hello, would you share your defined Mapper, looking for your reply
", "votes": 0, "replies": [] }, { "id": 1603996, "author_name": "Chan Kha Vu", "author_url": "", "post_date": "2021-12-03T02:30:52.027000", "content": "@zaopolearning hey, I added the mapper in this gist: https://gist.github.com/hav4ik/2d80b63dcab69651ad7bf1495053e35c - basically almost the same as the mapper in detectron documentation: https://detectron2.readthedocs.io/en/latest/tutorials/data_loading.html
\n(as you can see I use detectron2's transforms for geometric transforms like resizing and cropping, and albumentation for enhancing the image but…… augmentations doesn't seem to work here)
\nsorry for messy code lol :)
", "votes": 0, "replies": [] }, { "id": 1604006, "author_name": "yerongg", "author_url": "", "post_date": "2021-12-03T02:57:01.307000", "content": "thank you!
", "votes": 0, "replies": [] }, { "id": 1604012, "author_name": "", "author_url": "", "post_date": "2021-12-03T03:12:17.633000", "content": "", "votes": 0, "replies": [] }, { "id": 1606219, "author_name": "Shreya Sajal", "author_url": "", "post_date": "2021-12-04T22:37:35.390000", "content": "Hey @chankhavu ,thankyou for sharing this
\nI was trying to implement this and am getting the error in this part of the code:
\norig_img = cv2.imread(dataset_dict[\"file_name\"])
TypeError: list indices must be integers or slices, not str
\nMy dataset_dict is also a list of dicts (as you gave example of train_ds).
\nI don't know how to resolve this.
It doesn't seem to work when my images are different sizes.
", "votes": 0, "replies": [] } ] }, { "id": 1607980, "author_name": "shinning2021", "author_url": "", "post_date": "2021-12-06T07:01:15.433000", "content": "does anyone tried this method in mmdetection?
", "votes": 1, "replies": [] }, { "id": 1601005, "author_name": "Trushant Kalyanpur", "author_url": "", "post_date": "2021-11-30T23:18:55.440000", "content": "Thanks @mlneo07. Were you able to try this out? Looks like mmdet hasnt merged this feature yet
", "votes": 1, "replies": [ { "id": 1601346, "author_name": "Neo", "author_url": "", "post_date": "2021-12-01T07:23:05.787000", "content": "Yes, Slight improvement. I am using albumentations. MMDet is not merged yet but we can try @PIPELINES.register_module() for cusotmize dataset.
", "votes": 1, "replies": [] }, { "id": 1601374, "author_name": "Trushant Kalyanpur", "author_url": "", "post_date": "2021-12-01T08:12:38.030000", "content": "thanks @mlneo07 . Do you have an example of the custom pipeline file? I tried adding the decorator to the CopyPaste class but I think that is not the right approach.
Do you use copy-paste in MMDetection?
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1593998": "[Simple Copy-Paste is a Strong Data Augmentation Method\nfor Instance Segmentation](https://arxiv.org/pdf/2012.07177v2.pdf)\n\n\n\n![ ] (https://i.imgur.com/VBLFUdO.png)\n\nEdited::\n\nSimpleCopy Paste Augmentation Implementation in MMDetection\nhttps://github.com/open-mmlab/mmdetection/pull/6282\n`\ntrain_pipeline = [\n dict(type='SimpleCopyPaste', prob=0),\n dict(\n type='Resize',\n img_scale=[(1333, 640), (1333, 672), (1333, 704), (1333, 736),\n (1333, 768), (1333, 800)],\n multiscale_mode='value',\n keep_ratio=True),\n dict(type='RandomCrop', crop_size=(1024, 1024)),\n dict(type='RandomFlip', flip_ratio=0.5),\n dict(type='Normalize', **img_norm_cfg),\n dict(type='Pad', size_divisor=32),\n dict(type='DefaultFormatBundle'),\n dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks'])\n]`\n\n\nTo compose transforms with copy-paste augmentation with Albu::\n\nhttps://github.com/conradry/copy-paste-aug\n\n`import albumentations as A\nfrom albumentations.pytorch.transforms import ToTensorV2\nfrom copy_paste import CopyPaste\n\ntransform = A.Compose([\n A.RandomScale(scale_limit=(-0.9, 1), p=1), #LargeScaleJitter from scale of 0.1 to 2\n A.PadIfNeeded(256, 256, border_mode=0), #constant 0 border\n A.RandomCrop(256, 256),\n A.HorizontalFlip(p=0.5),\n CopyPaste(blend=True, sigma=1, pct_objects_paste=0.5, p=1)\n ], bbox_params=A.BboxParams(format=\"coco\")\n)`", "1601125": "I actually tried copy-paste augmentation (I'm using Detectron2, it is quite easy to implement). I see a slight improvement, but not that much.", "1607980": "does anyone tried this method in mmdetection?", "1601005": "Thanks @mlneo07. Were you able to try this out? Looks like mmdet hasnt merged this feature yet" } }