{
  "id": 413354,
  "title": "Does the model in the example need to be trained, or just use the pre-trained weights?",
  "url": "/competitions/image-matching-challenge-2023/discussion/413354",
  "author_name": "",
  "post_date": "2023-05-28T07:44:44.843046200Z",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Does the model in the example need to be trained, or just use the pre-trained weights? We looked at the top algorithms of the image matching challenge 2022, none of which have been trained, and they have achieved very high scores just by combining existing models. We tried to use the top ideas from last year, combining the loftr superglue and dkm models, but the score we got was less than 0.1. Could it be that the pairwise pairing of images is not accurate enough? Can the model and weights used for two image matching be used for this task simply without training?<br>\n🤓I am studying hard and hope to get your guidance！</p>",
  "messages": [
    {
      "id": "2277952",
      "postDate": "05/28/2023 07:44:44",
      "content": "<p>Does the model in the example need to be trained, or just use the pre-trained weights? We looked at the top algorithms of the image matching challenge 2022, none of which have been trained, and they have achieved very high scores just by combining existing models. We tried to use the top ideas from last year, combining the loftr superglue and dkm models, but the score we got was less than 0.1. Could it be that the pairwise pairing of images is not accurate enough? Can the model and weights used for two image matching be used for this task simply without training?<br>\n🤓I am studying hard and hope to get your guidance！</p>",
      "rawMarkdown": "Does the model in the example need to be trained, or just use the pre-trained weights? We looked at the top algorithms of the image matching challenge 2022, none of which have been trained, and they have achieved very high scores just by combining existing models. We tried to use the top ideas from last year, combining the loftr superglue and dkm models, but the score we got was less than 0.1. Could it be that the pairwise pairing of images is not accurate enough? Can the model and weights used for two image matching be used for this task simply without training?\n🤓I am studying hard and hope to get your guidance！",
      "votes": null
    },
    {
      "id": "2278056",
      "postDate": "05/28/2023 09:38:58",
      "content": "<p>I believe promising scores can be achieved even by using pre-trained weights. In the early stages of the competition, I attempted to concatenate the results of keynet, disk, and loftr in the example, and the score reached 0.25. Perhaps you can visualize your approach on images with minimal viewpoint changes to see if there are any issues with the combination method.</p>",
      "rawMarkdown": "I believe promising scores can be achieved even by using pre-trained weights. In the early stages of the competition, I attempted to concatenate the results of keynet, disk, and loftr in the example, and the score reached 0.25. Perhaps you can visualize your approach on images with minimal viewpoint changes to see if there are any issues with the combination method.",
      "votes": null
    },
    {
      "id": "2278070",
      "postDate": "05/28/2023 09:56:00",
      "content": "<p>Thank you very much for your reply! ❤️<br>\nKnowing the possibility of combining nets to get better results, we will check the code again for issues.<br>\nAlso, is it important to pair images in pairs? The efficientNet is used in the example. Which do you think is better to use a CNN network or an algorithm that uses some global descriptors?</p>",
      "rawMarkdown": "Thank you very much for your reply! ❤️\nKnowing the possibility of combining nets to get better results, we will check the code again for issues.\nAlso, is it important to pair images in pairs? The efficientNet is used in the example. Which do you think is better to use a CNN network or an algorithm that uses some global descriptors?",
      "votes": null
    },
    {
      "id": "2278268",
      "postDate": "05/28/2023 13:59:48",
      "content": "<p>I also combined the model in the example, but it ran out of time. Could you tell me how to deal with this problem? It has been bothering me for a long time.</p>",
      "rawMarkdown": "I also combined the model in the example, but it ran out of time. Could you tell me how to deal with this problem? It has been bothering me for a long time.",
      "votes": null
    },
    {
      "id": "2279057",
      "postDate": "05/29/2023 06:30:42",
      "content": "<blockquote>\n  <p>I also combined the model in the example, but it ran out of time. Could you tell me how to deal with this problem? It has been bothering me for a long time.<br>\n  fixing the long side rather than the short side in the example</p>\n</blockquote>",
      "rawMarkdown": "> I also combined the model in the example, but it ran out of time. Could you tell me how to deal with this problem? It has been bothering me for a long time.\nfixing the long side rather than the short side in the example",
      "votes": null
    },
    {
      "id": "2279062",
      "postDate": "05/29/2023 06:32:07",
      "content": "<p>Sorry, I haven't done much research on that</p>",
      "rawMarkdown": "Sorry, I haven't done much research on that",
      "votes": null
    },
    {
      "id": "2279172",
      "postDate": "05/29/2023 07:02:58",
      "content": "<p>😃Thanks anyway!</p>",
      "rawMarkdown": "😃Thanks anyway!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2278056,
      "author_name": "leon567",
      "author_url": "",
      "post_date": "05/28/2023 09:38:58",
      "content": "<p>I believe promising scores can be achieved even by using pre-trained weights. In the early stages of the competition, I attempted to concatenate the results of keynet, disk, and loftr in the example, and the score reached 0.25. Perhaps you can visualize your approach on images with minimal viewpoint changes to see if there are any issues with the combination method.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2278070,
          "author_name": "shuozhang2003",
          "author_url": "",
          "post_date": "05/28/2023 09:56:00",
          "content": "<p>Thank you very much for your reply! ❤️<br>\nKnowing the possibility of combining nets to get better results, we will check the code again for issues.<br>\nAlso, is it important to pair images in pairs? The efficientNet is used in the example. Which do you think is better to use a CNN network or an algorithm that uses some global descriptors?</p>",
          "votes": null,
          "replies": [
            {
              "id": 2279062,
              "author_name": "leon567",
              "author_url": "",
              "post_date": "05/29/2023 06:32:07",
              "content": "<p>Sorry, I haven't done much research on that</p>",
              "votes": null,
              "replies": [
                {
                  "id": 2279172,
                  "author_name": "shuozhang2003",
                  "author_url": "",
                  "post_date": "05/29/2023 07:02:58",
                  "content": "<p>😃Thanks anyway!</p>",
                  "votes": null,
                  "replies": []
                }
              ]
            }
          ]
        },
        {
          "id": 2278268,
          "author_name": "bent1e",
          "author_url": "",
          "post_date": "05/28/2023 13:59:48",
          "content": "<p>I also combined the model in the example, but it ran out of time. Could you tell me how to deal with this problem? It has been bothering me for a long time.</p>",
          "votes": null,
          "replies": [
            {
              "id": 2279057,
              "author_name": "leon567",
              "author_url": "",
              "post_date": "05/29/2023 06:30:42",
              "content": "<blockquote>\n  <p>I also combined the model in the example, but it ran out of time. Could you tell me how to deal with this problem? It has been bothering me for a long time.<br>\n  fixing the long side rather than the short side in the example</p>\n</blockquote>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2277952": "Does the model in the example need to be trained, or just use the pre-trained weights? We looked at the top algorithms of the image matching challenge 2022, none of which have been trained, and they have achieved very high scores just by combining existing models. We tried to use the top ideas from last year, combining the loftr superglue and dkm models, but the score we got was less than 0.1. Could it be that the pairwise pairing of images is not accurate enough? Can the model and weights used for two image matching be used for this task simply without training?\n🤓I am studying hard and hope to get your guidance！",
    "2278056": "I believe promising scores can be achieved even by using pre-trained weights. In the early stages of the competition, I attempted to concatenate the results of keynet, disk, and loftr in the example, and the score reached 0.25. Perhaps you can visualize your approach on images with minimal viewpoint changes to see if there are any issues with the combination method.",
    "2278070": "Thank you very much for your reply! ❤️\nKnowing the possibility of combining nets to get better results, we will check the code again for issues.\nAlso, is it important to pair images in pairs? The efficientNet is used in the example. Which do you think is better to use a CNN network or an algorithm that uses some global descriptors?",
    "2278268": "I also combined the model in the example, but it ran out of time. Could you tell me how to deal with this problem? It has been bothering me for a long time.",
    "2279057": "> I also combined the model in the example, but it ran out of time. Could you tell me how to deal with this problem? It has been bothering me for a long time.\nfixing the long side rather than the short side in the example",
    "2279062": "Sorry, I haven't done much research on that",
    "2279172": "😃Thanks anyway!"
  },
  "source": "meta"
}