{
  "id": 414392,
  "title": "LoFTR performs differently on Kaggle vs on my local machine",
  "url": "/competitions/image-matching-challenge-2023/discussion/414392",
  "author_name": "",
  "post_date": "2023-06-01T13:25:06.834503800Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I noticed that LoFTR produces different results on my local machine vs on Kaggle. I ran the exact same code and used the exact same model weights for LoFTR while doing validation on the same scene on my own computer and on Kaggle, but the results on Kaggle are always much worse than on my own computer (e.g., bike -- 0.09(Kaggle) vs 0.22(local), fountain -- between 0.35 to 0.65 (Kaggle) vs 0.91(local)) </p>\n<p>And this happens consistently only on LoFTR, I also tried KeyNetAffNetHardNet but it performs the same on both platforms. </p>\n<p>I wonder if anyone else has had this problem? I wonder what could be causing this, the only thing I can think of right now is my local pycolmap version is 0.4.0, which is different from 0.3.0 on Kaggle, but I don't think it would matter this much. Does anyone have any idea why?  </p>",
  "messages": [
    {
      "id": "2283741",
      "postDate": "06/01/2023 13:25:06",
      "content": "<p>I noticed that LoFTR produces different results on my local machine vs on Kaggle. I ran the exact same code and used the exact same model weights for LoFTR while doing validation on the same scene on my own computer and on Kaggle, but the results on Kaggle are always much worse than on my own computer (e.g., bike -- 0.09(Kaggle) vs 0.22(local), fountain -- between 0.35 to 0.65 (Kaggle) vs 0.91(local)) </p>\n<p>And this happens consistently only on LoFTR, I also tried KeyNetAffNetHardNet but it performs the same on both platforms. </p>\n<p>I wonder if anyone else has had this problem? I wonder what could be causing this, the only thing I can think of right now is my local pycolmap version is 0.4.0, which is different from 0.3.0 on Kaggle, but I don't think it would matter this much. Does anyone have any idea why?  </p>",
      "rawMarkdown": "I noticed that LoFTR produces different results on my local machine vs on Kaggle. I ran the exact same code and used the exact same model weights for LoFTR while doing validation on the same scene on my own computer and on Kaggle, but the results on Kaggle are always much worse than on my own computer (e.g., bike -- 0.09(Kaggle) vs 0.22(local), fountain -- between 0.35 to 0.65 (Kaggle) vs 0.91(local)) \n\nAnd this happens consistently only on LoFTR, I also tried KeyNetAffNetHardNet but it performs the same on both platforms. \n\nI wonder if anyone else has had this problem? I wonder what could be causing this, the only thing I can think of right now is my local pycolmap version is 0.4.0, which is different from 0.3.0 on Kaggle, but I don't think it would matter this much. Does anyone have any idea why?",
      "votes": null
    },
    {
      "id": "2283827",
      "postDate": "06/01/2023 14:23:19",
      "content": "<p>Maybe related to this <a href=\"https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/403568\" target=\"_blank\">https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/403568</a></p>",
      "rawMarkdown": "Maybe related to this https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/403568",
      "votes": null
    },
    {
      "id": "2283974",
      "postDate": "06/01/2023 16:30:23",
      "content": "<p>you can try install 0.4.0 to kaggle by downloading wheels, saving them as dataset and loading from them</p>",
      "rawMarkdown": "you can try install 0.4.0 to kaggle by downloading wheels, saving them as dataset and loading from them",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2283827,
      "author_name": "vostankovich",
      "author_url": "",
      "post_date": "06/01/2023 14:23:19",
      "content": "<p>Maybe related to this <a href=\"https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/403568\" target=\"_blank\">https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/403568</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2283974,
      "author_name": "oldufo",
      "author_url": "",
      "post_date": "06/01/2023 16:30:23",
      "content": "<p>you can try install 0.4.0 to kaggle by downloading wheels, saving them as dataset and loading from them</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2283741": "I noticed that LoFTR produces different results on my local machine vs on Kaggle. I ran the exact same code and used the exact same model weights for LoFTR while doing validation on the same scene on my own computer and on Kaggle, but the results on Kaggle are always much worse than on my own computer (e.g., bike -- 0.09(Kaggle) vs 0.22(local), fountain -- between 0.35 to 0.65 (Kaggle) vs 0.91(local)) \n\nAnd this happens consistently only on LoFTR, I also tried KeyNetAffNetHardNet but it performs the same on both platforms. \n\nI wonder if anyone else has had this problem? I wonder what could be causing this, the only thing I can think of right now is my local pycolmap version is 0.4.0, which is different from 0.3.0 on Kaggle, but I don't think it would matter this much. Does anyone have any idea why?",
    "2283827": "Maybe related to this https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/403568",
    "2283974": "you can try install 0.4.0 to kaggle by downloading wheels, saving them as dataset and loading from them"
  },
  "source": "meta"
}