{
  "id": 510596,
  "title": "17th Place Solution",
  "url": "/competitions/image-matching-challenge-2024/discussion/510596",
  "author_name": "",
  "post_date": "2024-06-06T18:55:43.093000",
  "votes": 11,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Firstly, I would like to express my gratitude to the Kaggle platform and the organizers of the Image Matching Challenge 2024.</p>\n<h3>Solution</h3>\n<p>Our solution refers to <a href=\"https://www.kaggle.com/code/maxchen303/imc2023-final-pub\" target=\"_blank\">the fifth place solution</a> from last year's competition, using keynet to detect key points in the image, Aliked for feature extraction, and AdaLAM matcher method. Using COLMAP for feature matching and incremental 3D reconstruction, including performing RANSAC algorithm to calculate the base matrix, performing beam adjustment to calculate the camera matrix, and rotating translation vector. Use multi process execution to accelerate processing and ensure program stability and efficiency.</p>\n<p>Set the number of features in KeyNetEffNetHardNet to 16000 and the minimum side length to 1200, which results in better performance. However, larger features may cause timeout and OOM. Although <a href=\"https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045\" target=\"_blank\">MAXCHEN303</a> mentioned in the scheme that setting \"force seed_mnn\" to True in the AdaLAM parameter section and increasing \"ransac_iters\" to 256 can improve the effectiveness of the scheme. However, after testing, setting the parameter 'force_seed_mnn' to False, 'ransac_iters' to 128, and' search_expansion 'to 64 seems to yield better results.</p>\n<p>We also attempted to set different resolutions for images in different scenes, but this method did not show significant results in LB. However, we found that this method displayed better results in PB. </p>\n<table>\n<thead>\n<tr>\n<th></th>\n<th>LB</th>\n<th>PB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Same resolution</td>\n<td>0.171938</td>\n<td>0.168047</td>\n</tr>\n<tr>\n<td>Different resolutions</td>\n<td>0.171806</td>\n<td>0.174287</td>\n</tr>\n</tbody>\n</table>\n<p><br></p>\n<p>This competition is greatly affected by randomness. We have repeatedly submitted the same code, but the scores are different each time. Our luck is not very good. The final submission solution was the two submissions with the highest public scores (version 2 and version 4) selected by the system by default, but their private scores were relatively low. The other two submissions with the same final solution had private scores that were very close to the gold medal scores. </p>\n<table>\n<thead>\n<tr>\n<th></th>\n<th>LB</th>\n<th>PB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Version 2 First submission</td>\n<td>0.180804</td>\n<td>0.175729</td>\n</tr>\n<tr>\n<td>Version 2 Second submission</td>\n<td>0.172519</td>\n<td>0.176941</td>\n</tr>\n<tr>\n<td>Version 4 First submission</td>\n<td>0.176453</td>\n<td>0.177361</td>\n</tr>\n<tr>\n<td>Version 4 Second submission</td>\n<td>0.176506</td>\n<td>0.174152</td>\n</tr>\n</tbody>\n</table>\n<h3>References</h3>\n<p><a href=\"https://www.kaggle.com/code/maxchen303/imc2023-final-pub\" target=\"_blank\">imc2023-final-pub</a><br>\n<a href=\"https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045\" target=\"_blank\">6th Place Solution - [ --- ]AffNetHardNet8 + AdaLAM</a></p>",
  "messages": [
    {
      "id": 2859038,
      "postDate": "2024-06-06T18:55:43.093Z",
      "content": "<p>Firstly, I would like to express my gratitude to the Kaggle platform and the organizers of the Image Matching Challenge 2024.</p>\n<h3>Solution</h3>\n<p>Our solution refers to <a href=\"https://www.kaggle.com/code/maxchen303/imc2023-final-pub\" target=\"_blank\">the fifth place solution</a> from last year's competition, using keynet to detect key points in the image, Aliked for feature extraction, and AdaLAM matcher method. Using COLMAP for feature matching and incremental 3D reconstruction, including performing RANSAC algorithm to calculate the base matrix, performing beam adjustment to calculate the camera matrix, and rotating translation vector. Use multi process execution to accelerate processing and ensure program stability and efficiency.</p>\n<p>Set the number of features in KeyNetEffNetHardNet to 16000 and the minimum side length to 1200, which results in better performance. However, larger features may cause timeout and OOM. Although <a href=\"https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045\" target=\"_blank\">MAXCHEN303</a> mentioned in the scheme that setting \"force seed_mnn\" to True in the AdaLAM parameter section and increasing \"ransac_iters\" to 256 can improve the effectiveness of the scheme. However, after testing, setting the parameter 'force_seed_mnn' to False, 'ransac_iters' to 128, and' search_expansion 'to 64 seems to yield better results.</p>\n<p>We also attempted to set different resolutions for images in different scenes, but this method did not show significant results in LB. However, we found that this method displayed better results in PB. </p>\n<table>\n<thead>\n<tr>\n<th></th>\n<th>LB</th>\n<th>PB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Same resolution</td>\n<td>0.171938</td>\n<td>0.168047</td>\n</tr>\n<tr>\n<td>Different resolutions</td>\n<td>0.171806</td>\n<td>0.174287</td>\n</tr>\n</tbody>\n</table>\n<p><br></p>\n<p>This competition is greatly affected by randomness. We have repeatedly submitted the same code, but the scores are different each time. Our luck is not very good. The final submission solution was the two submissions with the highest public scores (version 2 and version 4) selected by the system by default, but their private scores were relatively low. The other two submissions with the same final solution had private scores that were very close to the gold medal scores. </p>\n<table>\n<thead>\n<tr>\n<th></th>\n<th>LB</th>\n<th>PB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Version 2 First submission</td>\n<td>0.180804</td>\n<td>0.175729</td>\n</tr>\n<tr>\n<td>Version 2 Second submission</td>\n<td>0.172519</td>\n<td>0.176941</td>\n</tr>\n<tr>\n<td>Version 4 First submission</td>\n<td>0.176453</td>\n<td>0.177361</td>\n</tr>\n<tr>\n<td>Version 4 Second submission</td>\n<td>0.176506</td>\n<td>0.174152</td>\n</tr>\n</tbody>\n</table>\n<h3>References</h3>\n<p><a href=\"https://www.kaggle.com/code/maxchen303/imc2023-final-pub\" target=\"_blank\">imc2023-final-pub</a><br>\n<a href=\"https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045\" target=\"_blank\">6th Place Solution - [ --- ]AffNetHardNet8 + AdaLAM</a></p>",
      "rawMarkdown": "Firstly, I would like to express my gratitude to the Kaggle platform and the organizers of the Image Matching Challenge 2024.\n\n### Solution\nOur solution refers to [the fifth place solution](https://www.kaggle.com/code/maxchen303/imc2023-final-pub) from last year's competition, using keynet to detect key points in the image, Aliked for feature extraction, and AdaLAM matcher method. Using COLMAP for feature matching and incremental 3D reconstruction, including performing RANSAC algorithm to calculate the base matrix, performing beam adjustment to calculate the camera matrix, and rotating translation vector. Use multi process execution to accelerate processing and ensure program stability and efficiency.\n\nSet the number of features in KeyNetEffNetHardNet to 16000 and the minimum side length to 1200, which results in better performance. However, larger features may cause timeout and OOM. Although [MAXCHEN303](https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045) mentioned in the scheme that setting \"force seed_mnn\" to True in the AdaLAM parameter section and increasing \"ransac_iters\" to 256 can improve the effectiveness of the scheme. However, after testing, setting the parameter 'force_seed_mnn' to False, 'ransac_iters' to 128, and' search_expansion 'to 64 seems to yield better results.\n\nWe also attempted to set different resolutions for images in different scenes, but this method did not show significant results in LB. However, we found that this method displayed better results in PB. \n|       | LB       | PB       |\n|-------|----------|----------|\n| Same resolution | 0.171938 | 0.168047 |\n| Different resolutions | 0.171806 | 0.174287 |\n\n<br>\n\nThis competition is greatly affected by randomness. We have repeatedly submitted the same code, but the scores are different each time. Our luck is not very good. The final submission solution was the two submissions with the highest public scores (version 2 and version 4) selected by the system by default, but their private scores were relatively low. The other two submissions with the same final solution had private scores that were very close to the gold medal scores. \n|     | LB         | PB         |\n|--------|------------|------------|\n| Version 2 First submission | 0.180804   | 0.175729   |\n| Version 2 Second submission | 0.172519   | 0.176941   |\n| Version 4 First submission | 0.176453   | 0.177361   |\n| Version 4 Second submission | 0.176506   | 0.174152   |\n\n### References\n[imc2023-final-pub](https://www.kaggle.com/code/maxchen303/imc2023-final-pub)\n[6th Place Solution - [ --- ]AffNetHardNet8 + AdaLAM](https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045)",
      "votes": 11
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2859038": "Firstly, I would like to express my gratitude to the Kaggle platform and the organizers of the Image Matching Challenge 2024.\n\n### Solution\nOur solution refers to [the fifth place solution](https://www.kaggle.com/code/maxchen303/imc2023-final-pub) from last year's competition, using keynet to detect key points in the image, Aliked for feature extraction, and AdaLAM matcher method. Using COLMAP for feature matching and incremental 3D reconstruction, including performing RANSAC algorithm to calculate the base matrix, performing beam adjustment to calculate the camera matrix, and rotating translation vector. Use multi process execution to accelerate processing and ensure program stability and efficiency.\n\nSet the number of features in KeyNetEffNetHardNet to 16000 and the minimum side length to 1200, which results in better performance. However, larger features may cause timeout and OOM. Although [MAXCHEN303](https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045) mentioned in the scheme that setting \"force seed_mnn\" to True in the AdaLAM parameter section and increasing \"ransac_iters\" to 256 can improve the effectiveness of the scheme. However, after testing, setting the parameter 'force_seed_mnn' to False, 'ransac_iters' to 128, and' search_expansion 'to 64 seems to yield better results.\n\nWe also attempted to set different resolutions for images in different scenes, but this method did not show significant results in LB. However, we found that this method displayed better results in PB. \n|       | LB       | PB       |\n|-------|----------|----------|\n| Same resolution | 0.171938 | 0.168047 |\n| Different resolutions | 0.171806 | 0.174287 |\n\n<br>\n\nThis competition is greatly affected by randomness. We have repeatedly submitted the same code, but the scores are different each time. Our luck is not very good. The final submission solution was the two submissions with the highest public scores (version 2 and version 4) selected by the system by default, but their private scores were relatively low. The other two submissions with the same final solution had private scores that were very close to the gold medal scores. \n|     | LB         | PB         |\n|--------|------------|------------|\n| Version 2 First submission | 0.180804   | 0.175729   |\n| Version 2 Second submission | 0.172519   | 0.176941   |\n| Version 4 First submission | 0.176453   | 0.177361   |\n| Version 4 Second submission | 0.176506   | 0.174152   |\n\n### References\n[imc2023-final-pub](https://www.kaggle.com/code/maxchen303/imc2023-final-pub)\n[6th Place Solution - [ --- ]AffNetHardNet8 + AdaLAM](https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045)"
  }
}