In the training dataset, for each scene, there is an sfm folder which contains the file images.txt. In this file, the following data is present:
\n\n\nImage list with two lines of data per image:
\n
\n IMAGE_ID, QW, QX, QY, QZ, TX, TY, TZ, CAMERA_ID, NAME
\n POINTS2D[] as (X, Y, POINT3D_ID)
From this, I understand that QW, QX, QY, and QZ represent the rotation in quaternion (which I transformed into an R matrix), and TX, TY, and TZ represent the translation. However, in the file train_labels.csv, the rotation and translation matrices do not match with the aforementioned ones.
\nFurthermore, in the same file, there is a calibration matrix. So, for these training data, do we know the intrinsics, right? Because I read in another discussion that we don't know them.
\n@oldufo @fabiobellavia
In the training dataset, for each scene, there is an sfm folder which contains the file images.txt. In this file, the following data is present:
\n\n\nImage list with two lines of data per image:
\n
\n IMAGE_ID, QW, QX, QY, QZ, TX, TY, TZ, CAMERA_ID, NAME
\n POINTS2D[] as (X, Y, POINT3D_ID)
From this, I understand that QW, QX, QY, and QZ represent the rotation in quaternion (which I transformed into an R matrix), and TX, TY, and TZ represent the translation. However, in the file train_labels.csv, the rotation and translation matrices do not match with the aforementioned ones.
\nFurthermore, in the same file, there is a calibration matrix. So, for these training data, do we know the intrinsics, right? Because I read in another discussion that we don't know them.
\n@oldufo @fabiobellavia
\n\nFrom this, I understand that QW, QX, QY, and QZ represent the rotation in quaternion (which I transformed into an R matrix), and TX, TY, and TZ represent the translation. However, in the file train_labels.csv, the rotation and translation matrices do not match with the aforementioned ones.
\n
That is not true. If you read the reconstruction with colmap like:
\nimport pycolmap\nrec = pycolmap()\nimg=rec\nprint (())\nYou will get:
\narray([[- , , ],\n [-, , -],\n [-, -, ]])\nAnd for the train_labels.csv:
\nimport pandas as pd\ndf =pd()\nchurch_df = df==]\nimg111=church_df==]\n)\nYou will get
\n['-0.;0.734;0.1;-0.484;0.65;-0.653;-0.91;-0.395;0.56']\nWhich is exactly the same as rotation matrix.
\nRegarding the translation, though, the translation in colmap format is different from the train_labels.csv by some constant. The colmap translation is scale-less, whereas train_labels.csv gives you things in meters.
\nHowever, given that evaluation metric aligns the provided poses to the GT poses, this scale factor does not matter.
\n\nFurthermore, in the same file, there is a calibration matrix. So, for these training data, do we know the intrinsics, right? Because I read in another discussion that we don't know them.
\n
You know everything for the training data. For the test data you have only images.
", "rawMarkdown": ">From this, I understand that QW, QX, QY, and QZ represent the rotation in quaternion (which I transformed into an R matrix), and TX, TY, and TZ represent the translation. However, in the file train_labels.csv, the rotation and translation matrices do not match with the aforementioned ones.\n\nThat is not true. If you read the reconstruction with colmap like:\n\n```python3\nimport pycolmap\nrec = pycolmap.Reconstruction('path_to_church')\nimg=rec.images[111]\nprint (img.cam_from_world.rotation.matrix())\n```\n\nYou will get:\n\n```\narray([[-0.0180495 , 0.40600763, 0.91369142],\n [-0.47979086, 0.7982316 , -0.36417996],\n [-0.87719721, -0.44495406, 0.18039107]])\n```\n\nAnd for the train_labels.csv:\n```python3\nimport pandas as pd\ndf =pd.read_csv('train_labels.csv')\nchurch_df = df[df['dataset']=='church']\nimg111=church_df[church_df['image_name']=='00111.png']\nprint(img111.rotation_matrix.to_numpy())\n```\n\nYou will get \n```\n['-0.018049501107115562;0.40600763053442734;0.913691424638321;-0.47979085925197484;0.7982316005790165;-0.36417995993095653;-0.8771972109440591;-0.44495406030834395;0.1803910677585856']\n```\n\nWhich is exactly the same as rotation matrix. \nRegarding the translation, though, the translation in colmap format is different from the `train_labels.csv` by some constant. The colmap translation is scale-less, whereas `train_labels.csv` gives you things in meters. \nHowever, given that evaluation metric aligns the provided poses to the GT poses, this scale factor does not matter.\n\n\n>Furthermore, in the same file, there is a calibration matrix. So, for these training data, do we know the intrinsics, right? Because I read in another discussion that we don't know them.\n\nYou know everything for the training data. For the test data you have only images.", "votes": null }, { "id": "2772069", "postDate": "04/24/2024 14:11:42", "content": "Yes, you are right. I made a mistake in my calculations. Thank you for your answer! What are the units for the focal length and the principal points?
", "rawMarkdown": "Yes, you are right. I made a mistake in my calculations. Thank you for your answer! What are the units for the focal length and the principal points?", "votes": null }, { "id": "2773585", "postDate": "04/24/2024 19:42:36", "content": "They are pixels.
", "rawMarkdown": "They are pixels.", "votes": null } ], "comments": [ { "id": 2771510, "author_name": "oldufo", "author_url": "", "post_date": "04/24/2024 09:12:31", "content": "\n\nFrom this, I understand that QW, QX, QY, and QZ represent the rotation in quaternion (which I transformed into an R matrix), and TX, TY, and TZ represent the translation. However, in the file train_labels.csv, the rotation and translation matrices do not match with the aforementioned ones.
\n
That is not true. If you read the reconstruction with colmap like:
\nimport pycolmap\nrec = pycolmap()\nimg=rec\nprint (())\nYou will get:
\narray([[- , , ],\n [-, , -],\n [-, -, ]])\nAnd for the train_labels.csv:
\nimport pandas as pd\ndf =pd()\nchurch_df = df==]\nimg111=church_df==]\n)\nYou will get
\n['-0.;0.734;0.1;-0.484;0.65;-0.653;-0.91;-0.395;0.56']\nWhich is exactly the same as rotation matrix.
\nRegarding the translation, though, the translation in colmap format is different from the train_labels.csv by some constant. The colmap translation is scale-less, whereas train_labels.csv gives you things in meters.
\nHowever, given that evaluation metric aligns the provided poses to the GT poses, this scale factor does not matter.
\n\nFurthermore, in the same file, there is a calibration matrix. So, for these training data, do we know the intrinsics, right? Because I read in another discussion that we don't know them.
\n
You know everything for the training data. For the test data you have only images.
", "votes": null, "replies": [ { "id": 2772069, "author_name": "silavassiliou", "author_url": "", "post_date": "04/24/2024 14:11:42", "content": "Yes, you are right. I made a mistake in my calculations. Thank you for your answer! What are the units for the focal length and the principal points?
", "votes": null, "replies": [ { "id": 2773585, "author_name": "oldufo", "author_url": "", "post_date": "04/24/2024 19:42:36", "content": "They are pixels.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2769697": "In the training dataset, for each scene, there is an sfm folder which contains the file **images.txt**. In this file, the following data is present:\n>Image list with two lines of data per image:\nIMAGE_ID, QW, QX, QY, QZ, TX, TY, TZ, CAMERA_ID, NAME\nPOINTS2D[] as (X, Y, POINT3D_ID)\n\nFrom this, I understand that QW, QX, QY, and QZ represent the rotation in quaternion (which I transformed into an R matrix), and TX, TY, and TZ represent the translation. However, in the file **train_labels.csv,** the rotation and translation matrices do not match with the aforementioned ones.\n\nFurthermore, in the same file, there is a calibration matrix. So, for these training data, do we know the intrinsics, right? Because I read in another discussion that we don't know them. \n@oldufo @fabiobellavia", "2771510": ">From this, I understand that QW, QX, QY, and QZ represent the rotation in quaternion (which I transformed into an R matrix), and TX, TY, and TZ represent the translation. However, in the file train_labels.csv, the rotation and translation matrices do not match with the aforementioned ones.\n\nThat is not true. If you read the reconstruction with colmap like:\n\n```python3\nimport pycolmap\nrec = pycolmap.Reconstruction('path_to_church')\nimg=rec.images[111]\nprint (img.cam_from_world.rotation.matrix())\n```\n\nYou will get:\n\n```\narray([[-0.0180495 , 0.40600763, 0.91369142],\n [-0.47979086, 0.7982316 , -0.36417996],\n [-0.87719721, -0.44495406, 0.18039107]])\n```\n\nAnd for the train_labels.csv:\n```python3\nimport pandas as pd\ndf =pd.read_csv('train_labels.csv')\nchurch_df = df[df['dataset']=='church']\nimg111=church_df[church_df['image_name']=='00111.png']\nprint(img111.rotation_matrix.to_numpy())\n```\n\nYou will get \n```\n['-0.018049501107115562;0.40600763053442734;0.913691424638321;-0.47979085925197484;0.7982316005790165;-0.36417995993095653;-0.8771972109440591;-0.44495406030834395;0.1803910677585856']\n```\n\nWhich is exactly the same as rotation matrix. \nRegarding the translation, though, the translation in colmap format is different from the `train_labels.csv` by some constant. The colmap translation is scale-less, whereas `train_labels.csv` gives you things in meters. \nHowever, given that evaluation metric aligns the provided poses to the GT poses, this scale factor does not matter.\n\n\n>Furthermore, in the same file, there is a calibration matrix. So, for these training data, do we know the intrinsics, right? Because I read in another discussion that we don't know them.\n\nYou know everything for the training data. For the test data you have only images.", "2772069": "Yes, you are right. I made a mistake in my calculations. Thank you for your answer! What are the units for the focal length and the principal points?", "2773585": "They are pixels." }, "source": "meta" }