{ "id": 490984, "title": "What's the relation between the mean_reprojection_error and the final submission score ( mean Average Accuracy).", "url": "/competitions/image-matching-challenge-2024/discussion/490984", "author_name": "", "post_date": "2024-04-04T06:04:07.308572Z", "votes": 4, "comment_count": 4, "views": 0, "content": "

The initial base line solution had a

\n
\n

mean_reprojection_error of ~0.91 and mAA of 0.11.

\n
\n

After making some changes in the pipeline, i have a

\n
\n

mean_reprojection_error of ~0.72 and mAA of 0.01

\n
\n


\n
\nCan someone shed light on this please?

", "messages": [ { "id": "2734450", "postDate": "04/04/2024 06:04:07", "content": "

The initial base line solution had a

\n
\n

mean_reprojection_error of ~0.91 and mAA of 0.11.

\n
\n

After making some changes in the pipeline, i have a

\n
\n

mean_reprojection_error of ~0.72 and mAA of 0.01

\n
\n


\n
\nCan someone shed light on this please?

", "rawMarkdown": "The initial base line solution had a \n>mean_reprojection_error of ~0.91 and mAA of 0.11.\n\nAfter making some changes in the pipeline, i have a\n>mean_reprojection_error of ~0.72 and mAA of 0.01\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2697558%2F45f803d7a86905df086dc74939a5ecd0%2FScreenshot%202024-04-04%20112912.png?generation=1712210620121470&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2697558%2Fb5abfdc14803a20c1c0b7f48be56ab49%2FScreenshot%202024-04-04%20113008.png?generation=1712210632781744&alt=media)\nCan someone shed light on this please?", "votes": null }, { "id": "2734728", "postDate": "04/04/2024 09:19:04", "content": "

Hi, in general having a low reprojection error is good, and it is also expected since it is what you are minimizing in the bundle adjustment. This does not imply that the pose of the images is good, there is not direct relation between reproj error and mAA

", "rawMarkdown": "Hi, in general having a low reprojection error is good, and it is also expected since it is what you are minimizing in the bundle adjustment. This does not imply that the pose of the images is good, there is not direct relation between reproj error and mAA", "votes": null }, { "id": "2734745", "postDate": "04/04/2024 09:36:39", "content": "

You should treat reprojection error as “train set loss”, because Colmap directly optimizes it. And of course it can overfit, producing zero reprojection error, but completely wrong camera poses.

", "rawMarkdown": "You should treat reprojection error as “train set loss”, because Colmap directly optimizes it. And of course it can overfit, producing zero reprojection error, but completely wrong camera poses.", "votes": null }, { "id": "2734748", "postDate": "04/04/2024 09:40:49", "content": "

Thanks for the clarification. @lcmrll

", "rawMarkdown": "Thanks for the clarification. @lcmrll", "votes": null }, { "id": "2734749", "postDate": "04/04/2024 09:41:50", "content": "

I get it. Thanks @oldufo

", "rawMarkdown": "I get it. Thanks @oldufo", "votes": null } ], "comments": [ { "id": 2734728, "author_name": "lcmrll", "author_url": "", "post_date": "04/04/2024 09:19:04", "content": "

Hi, in general having a low reprojection error is good, and it is also expected since it is what you are minimizing in the bundle adjustment. This does not imply that the pose of the images is good, there is not direct relation between reproj error and mAA

", "votes": null, "replies": [ { "id": 2734748, "author_name": "adarshnanjaiya", "author_url": "", "post_date": "04/04/2024 09:40:49", "content": "

Thanks for the clarification. @lcmrll

", "votes": null, "replies": [] } ] }, { "id": 2734745, "author_name": "oldufo", "author_url": "", "post_date": "04/04/2024 09:36:39", "content": "

You should treat reprojection error as “train set loss”, because Colmap directly optimizes it. And of course it can overfit, producing zero reprojection error, but completely wrong camera poses.

", "votes": null, "replies": [ { "id": 2734749, "author_name": "adarshnanjaiya", "author_url": "", "post_date": "04/04/2024 09:41:50", "content": "

I get it. Thanks @oldufo

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2734450": "The initial base line solution had a \n>mean_reprojection_error of ~0.91 and mAA of 0.11.\n\nAfter making some changes in the pipeline, i have a\n>mean_reprojection_error of ~0.72 and mAA of 0.01\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2697558%2F45f803d7a86905df086dc74939a5ecd0%2FScreenshot%202024-04-04%20112912.png?generation=1712210620121470&alt=media)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2697558%2Fb5abfdc14803a20c1c0b7f48be56ab49%2FScreenshot%202024-04-04%20113008.png?generation=1712210632781744&alt=media)\nCan someone shed light on this please?", "2734728": "Hi, in general having a low reprojection error is good, and it is also expected since it is what you are minimizing in the bundle adjustment. This does not imply that the pose of the images is good, there is not direct relation between reproj error and mAA", "2734745": "You should treat reprojection error as “train set loss”, because Colmap directly optimizes it. And of course it can overfit, producing zero reprojection error, but completely wrong camera poses.", "2734748": "Thanks for the clarification. @lcmrll", "2734749": "I get it. Thanks @oldufo" }, "source": "meta" }