{
  "id": 512397,
  "title": "95th Place Solution:Effective Hyperparameter Tuning",
  "url": "/competitions/image-matching-challenge-2024/discussion/512397",
  "author_name": "HayatoFujihara",
  "post_date": "2024-06-15T02:01:56.993000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Thank you for hosting this competition! I learned a lot from it.</p>\n<p>Baseline:<a href=\"https://www.kaggle.com/code/motono0223/imc-2024-multi-models-pipeline?scriptVersionId=174818642\" target=\"_blank\">https://www.kaggle.com/code/motono0223/imc-2024-multi-models-pipeline?scriptVersionId=174818642</a><br>\n（Congratulations to the creator, motono0223, on becoming Competition Master!）</p>\n<p>Since this was my first image processing competition, I wasn't able to do anything special.However, looking at the top solutions from the previous competition, it was reported that scores could be improved by changing the image size or the number of keypoints.By adjusting some hyperparameters, the public score increased by 0.01, and the private score also increased by 0.0018 from the baseline public note.</p>\n<h3>Adjusted parameters</h3>\n<h4>params_aliked_lightglue</h4>\n<ul>\n<li>num_features = 8192 (experimented with many numbers from 512 to 12,000, but 8192 was the best fit.I also tried TTA with an image size of 4096 keypoints, but it was not able to beat 8192.)</li>\n<li>detection_threshold = 0.005</li>\n<li>min_matches = 100</li>\n<li>resize_to = 1500</li>\n</ul>\n<h4>def get_image_pairs_shortlist</h4>\n<ul>\n<li>sim_th = 0.08</li>\n</ul>",
  "messages": [
    {
      "id": 2872700,
      "postDate": "2024-06-15T02:01:56.993Z",
      "content": "<p>Thank you for hosting this competition! I learned a lot from it.</p>\n<p>Baseline:<a href=\"https://www.kaggle.com/code/motono0223/imc-2024-multi-models-pipeline?scriptVersionId=174818642\" target=\"_blank\">https://www.kaggle.com/code/motono0223/imc-2024-multi-models-pipeline?scriptVersionId=174818642</a><br>\n（Congratulations to the creator, motono0223, on becoming Competition Master!）</p>\n<p>Since this was my first image processing competition, I wasn't able to do anything special.However, looking at the top solutions from the previous competition, it was reported that scores could be improved by changing the image size or the number of keypoints.By adjusting some hyperparameters, the public score increased by 0.01, and the private score also increased by 0.0018 from the baseline public note.</p>\n<h3>Adjusted parameters</h3>\n<h4>params_aliked_lightglue</h4>\n<ul>\n<li>num_features = 8192 (experimented with many numbers from 512 to 12,000, but 8192 was the best fit.I also tried TTA with an image size of 4096 keypoints, but it was not able to beat 8192.)</li>\n<li>detection_threshold = 0.005</li>\n<li>min_matches = 100</li>\n<li>resize_to = 1500</li>\n</ul>\n<h4>def get_image_pairs_shortlist</h4>\n<ul>\n<li>sim_th = 0.08</li>\n</ul>",
      "rawMarkdown": "Thank you for hosting this competition! I learned a lot from it.\n\nBaseline:https://www.kaggle.com/code/motono0223/imc-2024-multi-models-pipeline?scriptVersionId=174818642\n（Congratulations to the creator, motono0223, on becoming Competition Master!）\n\nSince this was my first image processing competition, I wasn't able to do anything special.However, looking at the top solutions from the previous competition, it was reported that scores could be improved by changing the image size or the number of keypoints.By adjusting some hyperparameters, the public score increased by 0.01, and the private score also increased by 0.0018 from the baseline public note.\n\n\n### Adjusted parameters\n\n#### params_aliked_lightglue\n- num_features = 8192 (experimented with many numbers from 512 to 12,000, but 8192 was the best fit.I also tried TTA with an image size of 4096 keypoints, but it was not able to beat 8192.)\n- detection_threshold = 0.005\n- min_matches = 100\n- resize_to = 1500\n\n#### def get_image_pairs_shortlist\n- sim_th = 0.08",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2872700": "Thank you for hosting this competition! I learned a lot from it.\n\nBaseline:https://www.kaggle.com/code/motono0223/imc-2024-multi-models-pipeline?scriptVersionId=174818642\n（Congratulations to the creator, motono0223, on becoming Competition Master!）\n\nSince this was my first image processing competition, I wasn't able to do anything special.However, looking at the top solutions from the previous competition, it was reported that scores could be improved by changing the image size or the number of keypoints.By adjusting some hyperparameters, the public score increased by 0.01, and the private score also increased by 0.0018 from the baseline public note.\n\n\n### Adjusted parameters\n\n#### params_aliked_lightglue\n- num_features = 8192 (experimented with many numbers from 512 to 12,000, but 8192 was the best fit.I also tried TTA with an image size of 4096 keypoints, but it was not able to beat 8192.)\n- detection_threshold = 0.005\n- min_matches = 100\n- resize_to = 1500\n\n#### def get_image_pairs_shortlist\n- sim_th = 0.08"
  }
}