{
  "id": 575742,
  "title": "Should the resize parameter for ALIKED be modified through class variables?",
  "url": "/competitions/image-matching-challenge-2025/discussion/575742",
  "author_name": "",
  "post_date": "2025-04-30T15:48:37.798231100Z",
  "votes": 17,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I started participating from this year’s IMC, and I began by referring to <a href=\"https://www.kaggle.com/code/octaviograu/baseline-dinov2-aliked-lightglue\" target=\"_blank\">this baseline notebook</a>. (Thanks, <a href=\"https://www.kaggle.com/octaviograu\" target=\"_blank\">@octaviograu</a>!)</p>\n<p>I was puzzled that altering the <code>resize</code> argument in ALIKED didn’t change the results at all.</p>\n<pre><code>extractor = ALIKED(=num_features, =0.01, =resize_to)\n</code></pre>\n<p>After reviewing the implementation, I found that the resize parameter must be modified in the class variable <code>preprocess_conf</code> instead of being passed as the argument <a href=\"https://github.com/cvg/LightGlue/blob/edb2b838efb2ecfe3f88097c5fad9887d95aedad/lightglue/aliked.py#L631-L633\" target=\"_blank\">lightglue/aliked.py#L631-L633</a>.<br>\nTherefore, it needs to be modified as follows:</p>\n<pre><code>extractor = ALIKED(max_num_keypoints=num_features, detection_threshold=)\nextractor.preprocess_conf[] = resize_to\n</code></pre>\n<p>After changing resize from 1024 to 2048, my mAA score on the ETs benchmark increased from 44.23 % to 48.08 % !</p>",
  "messages": [
    {
      "id": "3190392",
      "postDate": "04/30/2025 15:48:37",
      "content": "<p>I started participating from this year’s IMC, and I began by referring to <a href=\"https://www.kaggle.com/code/octaviograu/baseline-dinov2-aliked-lightglue\" target=\"_blank\">this baseline notebook</a>. (Thanks, <a href=\"https://www.kaggle.com/octaviograu\" target=\"_blank\">@octaviograu</a>!)</p>\n<p>I was puzzled that altering the <code>resize</code> argument in ALIKED didn’t change the results at all.</p>\n<pre><code>extractor = ALIKED(=num_features, =0.01, =resize_to)\n</code></pre>\n<p>After reviewing the implementation, I found that the resize parameter must be modified in the class variable <code>preprocess_conf</code> instead of being passed as the argument <a href=\"https://github.com/cvg/LightGlue/blob/edb2b838efb2ecfe3f88097c5fad9887d95aedad/lightglue/aliked.py#L631-L633\" target=\"_blank\">lightglue/aliked.py#L631-L633</a>.<br>\nTherefore, it needs to be modified as follows:</p>\n<pre><code>extractor = ALIKED(max_num_keypoints=num_features, detection_threshold=)\nextractor.preprocess_conf[] = resize_to\n</code></pre>\n<p>After changing resize from 1024 to 2048, my mAA score on the ETs benchmark increased from 44.23 % to 48.08 % !</p>",
      "rawMarkdown": "I started participating from this year’s IMC, and I began by referring to [this baseline notebook](https://www.kaggle.com/code/octaviograu/baseline-dinov2-aliked-lightglue). (Thanks, @octaviograu!)\n\nI was puzzled that altering the `resize` argument in ALIKED didn’t change the results at all.\n```python:\nextractor = ALIKED(max_num_keypoints=num_features, detection_threshold=0.01, resize=resize_to)\n```\n\nAfter reviewing the implementation, I found that the resize parameter must be modified in the class variable `preprocess_conf` instead of being passed as the argument [lightglue/aliked.py#L631-L633](https://github.com/cvg/LightGlue/blob/edb2b838efb2ecfe3f88097c5fad9887d95aedad/lightglue/aliked.py#L631-L633).\nTherefore, it needs to be modified as follows:\n\n```python\nextractor = ALIKED(max_num_keypoints=num_features, detection_threshold=0.01)\nextractor.preprocess_conf[\"resize\"] = resize_to\n```\n\nAfter changing resize from 1024 to 2048, my mAA score on the ETs benchmark increased from 44.23 % to 48.08 % !",
      "votes": null
    },
    {
      "id": "3190479",
      "postDate": "04/30/2025 17:50:42",
      "content": "<p>Thank you for sharing this</p>",
      "rawMarkdown": "Thank you for sharing this",
      "votes": null
    },
    {
      "id": "3190788",
      "postDate": "05/01/2025 04:48:59",
      "content": "<p>Emm…, I would suggest to try:</p>\n<pre><code>=. --&gt; . --&gt; .\n</code></pre>",
      "rawMarkdown": "Emm..., I would suggest to try:\n```\ndetection_threshold=0.01 --> 0.05 --> 0.10\n```",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3190479,
      "author_name": "bishnuganguly",
      "author_url": "",
      "post_date": "04/30/2025 17:50:42",
      "content": "<p>Thank you for sharing this</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3190788,
      "author_name": "kenleungleung",
      "author_url": "",
      "post_date": "05/01/2025 04:48:59",
      "content": "<p>Emm…, I would suggest to try:</p>\n<pre><code>=. --&gt; . --&gt; .\n</code></pre>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3190392": "I started participating from this year’s IMC, and I began by referring to [this baseline notebook](https://www.kaggle.com/code/octaviograu/baseline-dinov2-aliked-lightglue). (Thanks, @octaviograu!)\n\nI was puzzled that altering the `resize` argument in ALIKED didn’t change the results at all.\n```python:\nextractor = ALIKED(max_num_keypoints=num_features, detection_threshold=0.01, resize=resize_to)\n```\n\nAfter reviewing the implementation, I found that the resize parameter must be modified in the class variable `preprocess_conf` instead of being passed as the argument [lightglue/aliked.py#L631-L633](https://github.com/cvg/LightGlue/blob/edb2b838efb2ecfe3f88097c5fad9887d95aedad/lightglue/aliked.py#L631-L633).\nTherefore, it needs to be modified as follows:\n\n```python\nextractor = ALIKED(max_num_keypoints=num_features, detection_threshold=0.01)\nextractor.preprocess_conf[\"resize\"] = resize_to\n```\n\nAfter changing resize from 1024 to 2048, my mAA score on the ETs benchmark increased from 44.23 % to 48.08 % !",
    "3190479": "Thank you for sharing this",
    "3190788": "Emm..., I would suggest to try:\n```\ndetection_threshold=0.01 --> 0.05 --> 0.10\n```"
  },
  "source": "meta"
}