{
  "id": 311370,
  "title": "Feature matching LoFTR - Kornia (Open CV - SURF comparision)",
  "url": "/competitions/herbarium-2022-fgvc9/discussion/311370",
  "author_name": "Remek Kinas",
  "post_date": "2022-03-06T14:20:17.998000",
  "votes": 6,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I am reading this competition and decided to share my latest discovery LoFTR. I know that my experiment is not directly connected with Herbarium but I think it could be helpful for you as well. I created LoFTR (feature matching) notebook for Whales competition which is currently active on Kaggle. Still I am thinking about joining here this is why I decided to take first step. </p>\n<p>I did SURF and LoFTR comparision …. Two observations:</p>\n<ul>\n<li>SURF and SIFT are not in Kaggle distribution -  algorithm is patented and is excluded in this configuration </li>\n<li>In my opinion SURF works worse then LoFTR on this dataset - it was difficult to find a match above 75% even on very similar fins (belonging to the same individual)</li>\n</ul>\n<p>See comparison:<br>\n<img src=\"https://i.ibb.co/smrJdDf/loftr01.jpg\" alt=\"Lo\"><br>\n<img src=\"https://i.ibb.co/VQ7qhGr/surf01.jpg\" alt=\"su\">  </p>\n<p>In notebook <strong>Whales feature matching LoFTR - Kornia</strong> I presented three things:</p>\n<ul>\n<li>LoFTR - Detector-Free Local Feature Matching with Transformers</li>\n<li>Kornia - absolutely fantastic library for Pytorch - Kornia is a differentiable library that allows classical computer vision to be integrated into deep learning models.</li>\n<li>Feature matching for images with background and without.</li>\n</ul>\n<p>Link to notebook: <a href=\"https://www.kaggle.com/remekkinas/whales-feature-matching-loftr-kornia\" target=\"_blank\">Whales feature matching LoFTR - Kornia</a> </p>",
  "messages": [
    {
      "id": 1713982,
      "postDate": "2022-03-06T14:20:18Z",
      "content": "<p>I am reading this competition and decided to share my latest discovery LoFTR. I know that my experiment is not directly connected with Herbarium but I think it could be helpful for you as well. I created LoFTR (feature matching) notebook for Whales competition which is currently active on Kaggle. Still I am thinking about joining here this is why I decided to take first step. </p>\n<p>I did SURF and LoFTR comparision …. Two observations:</p>\n<ul>\n<li>SURF and SIFT are not in Kaggle distribution -  algorithm is patented and is excluded in this configuration </li>\n<li>In my opinion SURF works worse then LoFTR on this dataset - it was difficult to find a match above 75% even on very similar fins (belonging to the same individual)</li>\n</ul>\n<p>See comparison:<br>\n<img src=\"https://i.ibb.co/smrJdDf/loftr01.jpg\" alt=\"Lo\"><br>\n<img src=\"https://i.ibb.co/VQ7qhGr/surf01.jpg\" alt=\"su\">  </p>\n<p>In notebook <strong>Whales feature matching LoFTR - Kornia</strong> I presented three things:</p>\n<ul>\n<li>LoFTR - Detector-Free Local Feature Matching with Transformers</li>\n<li>Kornia - absolutely fantastic library for Pytorch - Kornia is a differentiable library that allows classical computer vision to be integrated into deep learning models.</li>\n<li>Feature matching for images with background and without.</li>\n</ul>\n<p>Link to notebook: <a href=\"https://www.kaggle.com/remekkinas/whales-feature-matching-loftr-kornia\" target=\"_blank\">Whales feature matching LoFTR - Kornia</a> </p>",
      "rawMarkdown": "I am reading this competition and decided to share my latest discovery LoFTR. I know that my experiment is not directly connected with Herbarium but I think it could be helpful for you as well. I created LoFTR (feature matching) notebook for Whales competition which is currently active on Kaggle. Still I am thinking about joining here this is why I decided to take first step. \n\nI did SURF and LoFTR comparision .... Two observations:\n- SURF and SIFT are not in Kaggle distribution -  algorithm is patented and is excluded in this configuration \n- In my opinion SURF works worse then LoFTR on this dataset - it was difficult to find a match above 75% even on very similar fins (belonging to the same individual)\n\nSee comparison:\n![Lo](https://i.ibb.co/smrJdDf/loftr01.jpg)\n![su](https://i.ibb.co/VQ7qhGr/surf01.jpg)  \n\nIn notebook **Whales feature matching LoFTR - Kornia** I presented three things:\n- LoFTR - Detector-Free Local Feature Matching with Transformers\n- Kornia - absolutely fantastic library for Pytorch - Kornia is a differentiable library that allows classical computer vision to be integrated into deep learning models.\n- Feature matching for images with background and without.\n\nLink to notebook: [Whales feature matching LoFTR - Kornia](https://www.kaggle.com/remekkinas/whales-feature-matching-loftr-kornia) \n",
      "votes": 6
    },
    {
      "id": 1714024,
      "postDate": "2022-03-06T14:58:27.357Z",
      "content": "<p><a href=\"https://www.kaggle.com/RemekKinas\" target=\"_blank\">@RemekKinas</a> Thanks for the notes of caution with regard to SURF and SIFT and sharing your invaluable discovery of LoFTR. </p>",
      "rawMarkdown": "@RemekKinas Thanks for the notes of caution with regard to SURF and SIFT and sharing your invaluable discovery of LoFTR. ",
      "votes": 1,
      "replies": [
        {
          "id": 1714027,
          "postDate": "2022-03-06T15:00:30.753Z",
          "content": "<p><a href=\"https://www.kaggle.com/gianetan\" target=\"_blank\">@gianetan</a> thank you. For two reason this is good I think:</p>\n<ul>\n<li>supports GPU (Pytorch)</li>\n<li>it is really good in most cases</li>\n</ul>\n<p>I still working on improving this notebook (or solution). I thnink that on inference part of pipeline it could give us benefits.</p>",
          "rawMarkdown": "@gianetan thank you. For two reason this is good I think:\n- supports GPU (Pytorch)\n- it is really good in most cases\n\nI still working on improving this notebook (or solution). I thnink that on inference part of pipeline it could give us benefits."
        }
      ]
    },
    {
      "id": 1762329,
      "postDate": "2022-04-20T15:42:20.090Z",
      "content": "<p><a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> Thank you for always sharing your knowledge. <br>\n I was able to be in the top 8% of the HappyWhale competition.<br>\n As a beginner, I still have one question.<br>\nUnder what circumstances should I use LoFTR and Kornia libraries?</p>",
      "rawMarkdown": "@remekkinas Thank you for always sharing your knowledge. \n I was able to be in the top 8% of the HappyWhale competition.\n As a beginner, I still have one question.\nUnder what circumstances should I use LoFTR and Kornia libraries?"
    }
  ],
  "comments": [
    {
      "id": 1714024,
      "author_name": "Gianetan",
      "author_url": "",
      "post_date": "2022-03-06T14:58:27.357000",
      "content": "<p><a href=\"https://www.kaggle.com/RemekKinas\" target=\"_blank\">@RemekKinas</a> Thanks for the notes of caution with regard to SURF and SIFT and sharing your invaluable discovery of LoFTR. </p>",
      "votes": 1,
      "replies": [
        {
          "id": 1714027,
          "author_name": "Remek Kinas",
          "author_url": "",
          "post_date": "2022-03-06T15:00:30.753000",
          "content": "<p><a href=\"https://www.kaggle.com/gianetan\" target=\"_blank\">@gianetan</a> thank you. For two reason this is good I think:</p>\n<ul>\n<li>supports GPU (Pytorch)</li>\n<li>it is really good in most cases</li>\n</ul>\n<p>I still working on improving this notebook (or solution). I thnink that on inference part of pipeline it could give us benefits.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1762329,
      "author_name": "JaewooChoi",
      "author_url": "",
      "post_date": "2022-04-20T15:42:20.090000",
      "content": "<p><a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> Thank you for always sharing your knowledge. <br>\n I was able to be in the top 8% of the HappyWhale competition.<br>\n As a beginner, I still have one question.<br>\nUnder what circumstances should I use LoFTR and Kornia libraries?</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1713982": "I am reading this competition and decided to share my latest discovery LoFTR. I know that my experiment is not directly connected with Herbarium but I think it could be helpful for you as well. I created LoFTR (feature matching) notebook for Whales competition which is currently active on Kaggle. Still I am thinking about joining here this is why I decided to take first step. \n\nI did SURF and LoFTR comparision .... Two observations:\n- SURF and SIFT are not in Kaggle distribution -  algorithm is patented and is excluded in this configuration \n- In my opinion SURF works worse then LoFTR on this dataset - it was difficult to find a match above 75% even on very similar fins (belonging to the same individual)\n\nSee comparison:\n![Lo](https://i.ibb.co/smrJdDf/loftr01.jpg)\n![su](https://i.ibb.co/VQ7qhGr/surf01.jpg)  \n\nIn notebook **Whales feature matching LoFTR - Kornia** I presented three things:\n- LoFTR - Detector-Free Local Feature Matching with Transformers\n- Kornia - absolutely fantastic library for Pytorch - Kornia is a differentiable library that allows classical computer vision to be integrated into deep learning models.\n- Feature matching for images with background and without.\n\nLink to notebook: [Whales feature matching LoFTR - Kornia](https://www.kaggle.com/remekkinas/whales-feature-matching-loftr-kornia) \n",
    "1714024": "@RemekKinas Thanks for the notes of caution with regard to SURF and SIFT and sharing your invaluable discovery of LoFTR. ",
    "1762329": "@remekkinas Thank you for always sharing your knowledge. \n I was able to be in the top 8% of the HappyWhale competition.\n As a beginner, I still have one question.\nUnder what circumstances should I use LoFTR and Kornia libraries?"
  }
}