{
  "id": 320191,
  "title": "3rd solution - supplement to the post-processing section",
  "url": "/competitions/happy-whale-and-dolphin/discussion/320191",
  "author_name": "tik_boa",
  "post_date": "2022-04-20T12:40:53.838000",
  "votes": 10,
  "comment_count": 0,
  "views": 0,
  "content": "<p>First of all congratulations to all the winners 🎉， and thanks to kaggle for hosting such an interesting competition. </p>\n<p>A lot of post-processing strategies were tried in this competition，but most are disappointing.</p>\n<ul>\n<li>re-rank, accelerate with gpu and segmentation strategy. （not work）<ul>\n<li><a href=\"https://arxiv.org/pdf/1701.08398.pdf\" target=\"_blank\">https://arxiv.org/pdf/1701.08398.pdf</a> </li></ul></li>\n<li>dba, or neighbor blending. (not work)<ul>\n<li><a href=\"https://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf\" target=\"_blank\">https://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf</a></li></ul></li>\n<li>gnn-re-rank, based on graph convolution. (not work)<ul>\n<li><a href=\"https://arxiv.org/abs/2012.07620\" target=\"_blank\">https://arxiv.org/abs/2012.07620</a></li></ul></li>\n<li>some feature dimension processing strategies：PCA(part/whole), SVD(part/SVD), RMAC</li>\n<li>bayesian search weights<ul>\n<li>before 0.875, it can be greatly improved by searching</li>\n<li>but through repeated iterations, when the model reaches 0.880, the effect of search and concat is equivalent</li></ul></li>\n</ul>\n<p>Reference Code：</p>\n<ul>\n<li><a href=\"https://github.com/PyRetri/PyRetri\" target=\"_blank\">https://github.com/PyRetri/PyRetri</a></li>\n<li><a href=\"https://github.com/Xuanmeng-Zhang/gnn-re-ranking\" target=\"_blank\">https://github.com/Xuanmeng-Zhang/gnn-re-ranking</a></li>\n<li><a href=\"https://github.com/JDAI-CV/fast-reid\" target=\"_blank\">https://github.com/JDAI-CV/fast-reid</a></li>\n</ul>",
  "messages": [
    {
      "id": 1762103,
      "postDate": "2022-04-20T12:40:53.840Z",
      "content": "<p>First of all congratulations to all the winners 🎉， and thanks to kaggle for hosting such an interesting competition. </p>\n<p>A lot of post-processing strategies were tried in this competition，but most are disappointing.</p>\n<ul>\n<li>re-rank, accelerate with gpu and segmentation strategy. （not work）<ul>\n<li><a href=\"https://arxiv.org/pdf/1701.08398.pdf\" target=\"_blank\">https://arxiv.org/pdf/1701.08398.pdf</a> </li></ul></li>\n<li>dba, or neighbor blending. (not work)<ul>\n<li><a href=\"https://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf\" target=\"_blank\">https://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf</a></li></ul></li>\n<li>gnn-re-rank, based on graph convolution. (not work)<ul>\n<li><a href=\"https://arxiv.org/abs/2012.07620\" target=\"_blank\">https://arxiv.org/abs/2012.07620</a></li></ul></li>\n<li>some feature dimension processing strategies：PCA(part/whole), SVD(part/SVD), RMAC</li>\n<li>bayesian search weights<ul>\n<li>before 0.875, it can be greatly improved by searching</li>\n<li>but through repeated iterations, when the model reaches 0.880, the effect of search and concat is equivalent</li></ul></li>\n</ul>\n<p>Reference Code：</p>\n<ul>\n<li><a href=\"https://github.com/PyRetri/PyRetri\" target=\"_blank\">https://github.com/PyRetri/PyRetri</a></li>\n<li><a href=\"https://github.com/Xuanmeng-Zhang/gnn-re-ranking\" target=\"_blank\">https://github.com/Xuanmeng-Zhang/gnn-re-ranking</a></li>\n<li><a href=\"https://github.com/JDAI-CV/fast-reid\" target=\"_blank\">https://github.com/JDAI-CV/fast-reid</a></li>\n</ul>",
      "rawMarkdown": "First of all congratulations to all the winners 🎉， and thanks to kaggle for hosting such an interesting competition. \n\nA lot of post-processing strategies were tried in this competition，but most are disappointing.\n* re-rank, accelerate with gpu and segmentation strategy. （not work）\n * https://arxiv.org/pdf/1701.08398.pdf \n*  dba, or neighbor blending. (not work)\n * https://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf\n* gnn-re-rank, based on graph convolution. (not work)\n * https://arxiv.org/abs/2012.07620\n* some feature dimension processing strategies：PCA(part/whole), SVD(part/SVD), RMAC\n*  bayesian search weights\n * before 0.875, it can be greatly improved by searching\n * but through repeated iterations, when the model reaches 0.880, the effect of search and concat is equivalent\n\nReference Code：\n* https://github.com/PyRetri/PyRetri\n* https://github.com/Xuanmeng-Zhang/gnn-re-ranking\n* https://github.com/JDAI-CV/fast-reid",
      "votes": 10
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1762103": "First of all congratulations to all the winners 🎉， and thanks to kaggle for hosting such an interesting competition. \n\nA lot of post-processing strategies were tried in this competition，but most are disappointing.\n* re-rank, accelerate with gpu and segmentation strategy. （not work）\n * https://arxiv.org/pdf/1701.08398.pdf \n*  dba, or neighbor blending. (not work)\n * https://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf\n* gnn-re-rank, based on graph convolution. (not work)\n * https://arxiv.org/abs/2012.07620\n* some feature dimension processing strategies：PCA(part/whole), SVD(part/SVD), RMAC\n*  bayesian search weights\n * before 0.875, it can be greatly improved by searching\n * but through repeated iterations, when the model reaches 0.880, the effect of search and concat is equivalent\n\nReference Code：\n* https://github.com/PyRetri/PyRetri\n* https://github.com/Xuanmeng-Zhang/gnn-re-ranking\n* https://github.com/JDAI-CV/fast-reid"
  }
}