{
  "id": 414465,
  "title": "How to deal with scenes with a lot of camera rotation but little camera translation (like Cyprus)?",
  "url": "/competitions/image-matching-challenge-2023/discussion/414465",
  "author_name": "",
  "post_date": "2023-06-01T18:37:20.618988200Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I noticed some of the scenes in the training dataset (Cyprus and Dioscuri somewhat) have a lot of camera rotations but little camera translation, which made me not able to get any decent results at all (especially Cyprus Cyprus ). I think in order to improve, it would require fitting homography to obtain the accurate camera poses (both translation and rotation). I suppose fitting fundamental matrices can still get you decent rotation matrix but the translation vector could be completely wrong (which makes the results useless). I wonder if there's a good way of dealing with this scenario? </p>\n<p>I think the most naive way would be to run Colmap twice using fundamental matrix and homography on each scene and inspect the errors somehow to decide which model to pick? But it would be indeed very slow and I am not sure if it can work practically. How do you guys deal with these scenes? Or maybe they don't make a big difference in the final score?</p>",
  "messages": [
    {
      "id": "2284131",
      "postDate": "06/01/2023 18:37:20",
      "content": "<p>I noticed some of the scenes in the training dataset (Cyprus and Dioscuri somewhat) have a lot of camera rotations but little camera translation, which made me not able to get any decent results at all (especially Cyprus Cyprus ). I think in order to improve, it would require fitting homography to obtain the accurate camera poses (both translation and rotation). I suppose fitting fundamental matrices can still get you decent rotation matrix but the translation vector could be completely wrong (which makes the results useless). I wonder if there's a good way of dealing with this scenario? </p>\n<p>I think the most naive way would be to run Colmap twice using fundamental matrix and homography on each scene and inspect the errors somehow to decide which model to pick? But it would be indeed very slow and I am not sure if it can work practically. How do you guys deal with these scenes? Or maybe they don't make a big difference in the final score?</p>",
      "rawMarkdown": "I noticed some of the scenes in the training dataset (Cyprus and Dioscuri somewhat) have a lot of camera rotations but little camera translation, which made me not able to get any decent results at all (especially Cyprus Cyprus ). I think in order to improve, it would require fitting homography to obtain the accurate camera poses (both translation and rotation). I suppose fitting fundamental matrices can still get you decent rotation matrix but the translation vector could be completely wrong (which makes the results useless). I wonder if there's a good way of dealing with this scenario? \n\nI think the most naive way would be to run Colmap twice using fundamental matrix and homography on each scene and inspect the errors somehow to decide which model to pick? But it would be indeed very slow and I am not sure if it can work practically. How do you guys deal with these scenes? Or maybe they don't make a big difference in the final score?",
      "votes": null
    },
    {
      "id": "2285344",
      "postDate": "06/02/2023 16:26:16",
      "content": "<p>I only know that this type of data accounts for a large proportion of the final score. I have noticed that the challenges you are currently facing and the final score is similar to my team's. If you can solve this problem, the final score will be improved</p>",
      "rawMarkdown": "I only know that this type of data accounts for a large proportion of the final score. I have noticed that the challenges you are currently facing and the final score is similar to my team's. If you can solve this problem, the final score will be improved",
      "votes": null
    },
    {
      "id": "2285373",
      "postDate": "06/02/2023 16:46:45",
      "content": "<p>How do you know this type of data accounts for a large proportion of the final score? </p>",
      "rawMarkdown": "How do you know this type of data accounts for a large proportion of the final score?",
      "votes": null
    },
    {
      "id": "2287484",
      "postDate": "06/04/2023 14:16:33",
      "content": "<p>t seems that you have solved the problem. Congratulations!</p>",
      "rawMarkdown": "t seems that you have solved the problem. Congratulations!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2285344,
      "author_name": "bent1e",
      "author_url": "",
      "post_date": "06/02/2023 16:26:16",
      "content": "<p>I only know that this type of data accounts for a large proportion of the final score. I have noticed that the challenges you are currently facing and the final score is similar to my team's. If you can solve this problem, the final score will be improved</p>",
      "votes": null,
      "replies": [
        {
          "id": 2285373,
          "author_name": "anonymousyuxiang",
          "author_url": "",
          "post_date": "06/02/2023 16:46:45",
          "content": "<p>How do you know this type of data accounts for a large proportion of the final score? </p>",
          "votes": null,
          "replies": [
            {
              "id": 2287484,
              "author_name": "bent1e",
              "author_url": "",
              "post_date": "06/04/2023 14:16:33",
              "content": "<p>t seems that you have solved the problem. Congratulations!</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2284131": "I noticed some of the scenes in the training dataset (Cyprus and Dioscuri somewhat) have a lot of camera rotations but little camera translation, which made me not able to get any decent results at all (especially Cyprus Cyprus ). I think in order to improve, it would require fitting homography to obtain the accurate camera poses (both translation and rotation). I suppose fitting fundamental matrices can still get you decent rotation matrix but the translation vector could be completely wrong (which makes the results useless). I wonder if there's a good way of dealing with this scenario? \n\nI think the most naive way would be to run Colmap twice using fundamental matrix and homography on each scene and inspect the errors somehow to decide which model to pick? But it would be indeed very slow and I am not sure if it can work practically. How do you guys deal with these scenes? Or maybe they don't make a big difference in the final score?",
    "2285344": "I only know that this type of data accounts for a large proportion of the final score. I have noticed that the challenges you are currently facing and the final score is similar to my team's. If you can solve this problem, the final score will be improved",
    "2285373": "How do you know this type of data accounts for a large proportion of the final score?",
    "2287484": "t seems that you have solved the problem. Congratulations!"
  },
  "source": "meta"
}