We were given specific image pairs previous year and we were supposed to match them. If I understood correctly, this time we have to match all image pairs in a scene but there is only one image on each row in the submission file. How does the submission file work? @eduardtrulls @oldufo
", "messages": [ { "id": "2222568", "postDate": "04/15/2023 11:17:15", "content": "We were given specific image pairs previous year and we were supposed to match them. If I understood correctly, this time we have to match all image pairs in a scene but there is only one image on each row in the submission file. How does the submission file work? @eduardtrulls @oldufo
", "rawMarkdown": "We were given specific image pairs previous year and we were supposed to match them. If I understood correctly, this time we have to match all image pairs in a scene but there is only one image on each row in the submission file. How does the submission file work? @eduardtrulls @oldufo", "votes": null }, { "id": "2222581", "postDate": "04/15/2023 11:36:02", "content": "Basically, you have to match all pairs in the scene and then by SfM (bundle adjstument) retrieve consistent camera poses within the whole scene (this is the classic approach, but recent ones are welcome too). See for instance the multiview path here: https://www.cs.ubc.ca/research/image-matching-challenge/2021/benchmark/ .
\nYou can inspect / play with the submission example to understand effectively how the submission works: https://www.kaggle.com/code/eduardtrulls/imc-2023-submission-example
", "rawMarkdown": "Basically, you have to match all pairs in the scene and then by SfM (bundle adjstument) retrieve consistent camera poses within the whole scene (this is the classic approach, but recent ones are welcome too). See for instance the multiview path here: https://www.cs.ubc.ca/research/image-matching-challenge/2021/benchmark/ .\n\nYou can inspect / play with the submission example to understand effectively how the submission works: https://www.kaggle.com/code/eduardtrulls/imc-2023-submission-example", "votes": null }, { "id": "2222608", "postDate": "04/15/2023 12:08:14", "content": "Thanks for the fast reply. Is task 2 in 2021 competition exactly same as this competition's task?
", "rawMarkdown": "Thanks for the fast reply. Is task 2 in 2021 competition exactly same as this competition's task?", "votes": null }, { "id": "2222621", "postDate": "04/15/2023 12:25:11", "content": "Not exactly, previously the participants stopped after providing the pairwise keypoint matches, now they must set-up / adjust the whole SfM pipeline until they get he camera poses. The submission example uses COLMAP to get the poses but any other methods can be used.
", "rawMarkdown": "Not exactly, previously the participants stopped after providing the pairwise keypoint matches, now they must set-up / adjust the whole SfM pipeline until they get he camera poses. The submission example uses COLMAP to get the poses but any other methods can be used.", "votes": null }, { "id": "2222632", "postDate": "04/15/2023 12:40:58", "content": "Aren't poses from the previous year are actually pose of image 2 relative to image 1? What does the pose COLMAP outputs represent?
", "rawMarkdown": "Aren't poses from the previous year are actually pose of image 2 relative to image 1? What does the pose COLMAP outputs represent?", "votes": null }, { "id": "2222656", "postDate": "04/15/2023 13:02:42", "content": "For COLMAP format please refer to https://colmap.github.io/format.html , specifically \"images.txt\" and compare with the evaluation example: https://www.kaggle.com/code/eduardtrulls/imc2023-evaluation . Checking the whole COLMAP documentation and the related references can be useful to you.
\nSfM camera poses (rotation matrices /quaternions plus corresponding translation vectors) are ALL with respect to a COMMON reference system. Moreover, translations vectors are ALL up to a COMMON scale factor. Concatenating camera rigid transformations in SfM is OVERALL CONSISTENT.
", "rawMarkdown": "For COLMAP format please refer to https://colmap.github.io/format.html , specifically \"images.txt\" and compare with the evaluation example: https://www.kaggle.com/code/eduardtrulls/imc2023-evaluation . Checking the whole COLMAP documentation and the related references can be useful to you.\n\nSfM camera poses (rotation matrices /quaternions plus corresponding translation vectors) are ALL with respect to a COMMON reference system. Moreover, translations vectors are ALL up to a COMMON scale factor. Concatenating camera rigid transformations in SfM is OVERALL CONSISTENT.", "votes": null }, { "id": "2222944", "postDate": "04/15/2023 16:55:28", "content": "So after feeding all relative image pair poses to COLMAP, we can get a camera pose with respect to a common reference system for each image. COLMAP inherently allows that system so we don't worry about normalizing, re-scaling and etc. It's kinda similar to vector space.
", "rawMarkdown": "So after feeding all relative image pair poses to COLMAP, we can get a camera pose with respect to a common reference system for each image. COLMAP inherently allows that system so we don't worry about normalizing, re-scaling and etc. It's kinda similar to vector space.", "votes": null }, { "id": "2222948", "postDate": "04/15/2023 17:00:29", "content": "True, but COLMAP can fail to register some cameras / images !
", "rawMarkdown": "True, but COLMAP can fail to register some cameras / images !", "votes": null }, { "id": "2222955", "postDate": "04/15/2023 17:06:06", "content": "Right, but does COLMAP generalize? If we provide a sample of relative image pair poses, can it register an unseen image?
", "rawMarkdown": "Right, but does COLMAP generalize? If we provide a sample of relative image pair poses, can it register an unseen image?", "votes": null }, { "id": "2222972", "postDate": "04/15/2023 17:13:36", "content": "Play with COLMAP using the GUI, it's much better than asking questions :)
", "rawMarkdown": "Play with COLMAP using the GUI, it's much better than asking questions :)", "votes": null }, { "id": "2222978", "postDate": "04/15/2023 17:17:01", "content": "Yeah, you are right. Thanks for the help :D
", "rawMarkdown": "Yeah, you are right. Thanks for the help :D", "votes": null } ], "comments": [ { "id": 2222581, "author_name": "fabiobellavia", "author_url": "", "post_date": "04/15/2023 11:36:02", "content": "Basically, you have to match all pairs in the scene and then by SfM (bundle adjstument) retrieve consistent camera poses within the whole scene (this is the classic approach, but recent ones are welcome too). See for instance the multiview path here: https://www.cs.ubc.ca/research/image-matching-challenge/2021/benchmark/ .
\nYou can inspect / play with the submission example to understand effectively how the submission works: https://www.kaggle.com/code/eduardtrulls/imc-2023-submission-example
", "votes": null, "replies": [ { "id": 2222608, "author_name": "gunesevitan", "author_url": "", "post_date": "04/15/2023 12:08:14", "content": "Thanks for the fast reply. Is task 2 in 2021 competition exactly same as this competition's task?
", "votes": null, "replies": [ { "id": 2222621, "author_name": "fabiobellavia", "author_url": "", "post_date": "04/15/2023 12:25:11", "content": "Not exactly, previously the participants stopped after providing the pairwise keypoint matches, now they must set-up / adjust the whole SfM pipeline until they get he camera poses. The submission example uses COLMAP to get the poses but any other methods can be used.
", "votes": null, "replies": [ { "id": 2222632, "author_name": "gunesevitan", "author_url": "", "post_date": "04/15/2023 12:40:58", "content": "Aren't poses from the previous year are actually pose of image 2 relative to image 1? What does the pose COLMAP outputs represent?
", "votes": null, "replies": [ { "id": 2222656, "author_name": "fabiobellavia", "author_url": "", "post_date": "04/15/2023 13:02:42", "content": "For COLMAP format please refer to https://colmap.github.io/format.html , specifically \"images.txt\" and compare with the evaluation example: https://www.kaggle.com/code/eduardtrulls/imc2023-evaluation . Checking the whole COLMAP documentation and the related references can be useful to you.
\nSfM camera poses (rotation matrices /quaternions plus corresponding translation vectors) are ALL with respect to a COMMON reference system. Moreover, translations vectors are ALL up to a COMMON scale factor. Concatenating camera rigid transformations in SfM is OVERALL CONSISTENT.
", "votes": null, "replies": [ { "id": 2222944, "author_name": "gunesevitan", "author_url": "", "post_date": "04/15/2023 16:55:28", "content": "So after feeding all relative image pair poses to COLMAP, we can get a camera pose with respect to a common reference system for each image. COLMAP inherently allows that system so we don't worry about normalizing, re-scaling and etc. It's kinda similar to vector space.
", "votes": null, "replies": [ { "id": 2222948, "author_name": "fabiobellavia", "author_url": "", "post_date": "04/15/2023 17:00:29", "content": "True, but COLMAP can fail to register some cameras / images !
", "votes": null, "replies": [ { "id": 2222955, "author_name": "gunesevitan", "author_url": "", "post_date": "04/15/2023 17:06:06", "content": "Right, but does COLMAP generalize? If we provide a sample of relative image pair poses, can it register an unseen image?
", "votes": null, "replies": [ { "id": 2222972, "author_name": "fabiobellavia", "author_url": "", "post_date": "04/15/2023 17:13:36", "content": "Play with COLMAP using the GUI, it's much better than asking questions :)
", "votes": null, "replies": [ { "id": 2222978, "author_name": "gunesevitan", "author_url": "", "post_date": "04/15/2023 17:17:01", "content": "Yeah, you are right. Thanks for the help :D
", "votes": null, "replies": [] } ] } ] } ] } ] } ] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "2222568": "We were given specific image pairs previous year and we were supposed to match them. If I understood correctly, this time we have to match all image pairs in a scene but there is only one image on each row in the submission file. How does the submission file work? @eduardtrulls @oldufo", "2222581": "Basically, you have to match all pairs in the scene and then by SfM (bundle adjstument) retrieve consistent camera poses within the whole scene (this is the classic approach, but recent ones are welcome too). See for instance the multiview path here: https://www.cs.ubc.ca/research/image-matching-challenge/2021/benchmark/ .\n\nYou can inspect / play with the submission example to understand effectively how the submission works: https://www.kaggle.com/code/eduardtrulls/imc-2023-submission-example", "2222608": "Thanks for the fast reply. Is task 2 in 2021 competition exactly same as this competition's task?", "2222621": "Not exactly, previously the participants stopped after providing the pairwise keypoint matches, now they must set-up / adjust the whole SfM pipeline until they get he camera poses. The submission example uses COLMAP to get the poses but any other methods can be used.", "2222632": "Aren't poses from the previous year are actually pose of image 2 relative to image 1? What does the pose COLMAP outputs represent?", "2222656": "For COLMAP format please refer to https://colmap.github.io/format.html , specifically \"images.txt\" and compare with the evaluation example: https://www.kaggle.com/code/eduardtrulls/imc2023-evaluation . Checking the whole COLMAP documentation and the related references can be useful to you.\n\nSfM camera poses (rotation matrices /quaternions plus corresponding translation vectors) are ALL with respect to a COMMON reference system. Moreover, translations vectors are ALL up to a COMMON scale factor. Concatenating camera rigid transformations in SfM is OVERALL CONSISTENT.", "2222944": "So after feeding all relative image pair poses to COLMAP, we can get a camera pose with respect to a common reference system for each image. COLMAP inherently allows that system so we don't worry about normalizing, re-scaling and etc. It's kinda similar to vector space.", "2222948": "True, but COLMAP can fail to register some cameras / images !", "2222955": "Right, but does COLMAP generalize? If we provide a sample of relative image pair poses, can it register an unseen image?", "2222972": "Play with COLMAP using the GUI, it's much better than asking questions :)", "2222978": "Yeah, you are right. Thanks for the help :D" }, "source": "meta" }