{"id":8090,"date":"2022-09-22T17:18:00","date_gmt":"2022-09-22T09:18:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=8090"},"modified":"2022-11-25T10:29:55","modified_gmt":"2022-11-25T02:29:55","slug":"pointrend","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/09\/22\/pointrend\/","title":{"rendered":"PointRend –\u56fe\u50cf\u7ec6\u9897\u7c92\u5206\u5272"},"content":{"rendered":"\n\n\n
\"title\"
https:\/\/arxiv.org\/abs\/1912.08193<\/figcaption><\/figure>\n\n\n\n

\u8bba\u6587\u5730\u5740\uff1a https:\/\/arxiv.org\/abs\/1912.08193<\/a><\/strong><\/p>\n\n\n\n

gitlab: https:\/\/github.com\/zsef123\/PointRend-PyTorch<\/a><\/strong><\/p>\n\n\n\n

\u5b58\u5728\u7684\u95ee\u9898<\/strong><\/h2>\n\n\n\n

\u5728\u76ee\u524d\u7684\u8bed\u4e49\u5206\u5272\u7f51\u7edc\u4e2d\u5b58\u5728\u7684\u95ee\u9898\u4e3b\u8981\u6709\u8fc7\u91c7\u6837\u548c\u73b0\u91c7\u6837\u3002<\/p>\n\n\n\n

1.\u8fc7\u91c7\u6837\uff08 oversample \uff09\uff1a\u5bf9\u4e8e\u56fe\u7247\u4e2d\u4f4e\u9891\u533a\u57df\uff08 \u5c5e\u4e8e\u540c\u4e00\u4e2a\u7269\u4f53 \uff09\uff0c\u6ca1\u5fc5\u8981\u4f7f\u7528 \u592a\u591a\u7684\u91c7\u6837\u70b9\uff0c\u5374\u4f7f\u7528\u592a\u591a\u91c7\u6837\u70b9\u9020\u6210\u8fc7\u91c7\u6837\uff1b<\/p>\n\n\n\n

2.\u6b20\u91c7\u6837\uff08 undersample \uff09 \uff1a\u5bf9\u4e8e\u56fe\u7247\u4e2d\u9ad8\u9891\u533a\u57df\uff08 \u9760\u8fd1\u7269\u4f53\u8fb9\u754c \uff09\uff0c\u5982\u679c\u8fd9\u4e9b\u533a\u57df\u7684\u91c7\u6837\u8fc7\u4e8e\u7a00\u758f\uff0c\u5bfc\u81f4\u5206\u5272\u51fa\u7684\u8fb9\u754c\u8fc7\u4e8e\u5e73\u6ed1\uff0c\u4e0d\u5927\u771f\u5b9e<\/p>\n\n\n\n

\u6587\u7ae0\u8981\u89e3\u51b3\u7684\u95ee\u9898\u662f\u5728\u5b9e\u4f8b\u5206\u5272\u4efb\u52a1\u4e2d\u8fb9\u7f18\u4e0d\u591f\u7cbe\u7ec6\u7684\u95ee\u9898\u3002\u4ee5MaskRCNN\u4e3e\u4f8b\uff0c\u7531\u4e8e\u8ba1\u7b97\u91cf\u548c\u663e\u5b58\u7684\u539f\u56e0\uff0c\u5bf9\u4e8e\u6bcf\u4e00\u4e2aROIAlign\u4e4b\u540e\u7684proposal\u6211\u4eec\u4e00\u822c\u53ea\u4f1aupsample\u523028*28\u7684\u5206\u8fa8\u7387\u8f93\u51famask\u3002\u8fd9\u5bf9\u4e8e\u7edd\u5927\u591a\u6570\u7269\u4f53\u663e\u7136\u662f\u4e0d\u591f\u7684\u3002\u5982\u679c\u60f3\u5f97\u5230\u50cf\u7d20\u7ea7\u522b\u7684\u7cbe\u5ea6\uff0c\u6211\u4eec\u4e0d\u5f97\u4e0d\u4ed8\u51fa\u66f4\u5927\u7684\u8ba1\u7b97\u548c\u5b58\u50a8\u4ee3\u4ef7\u3002\u90a3\u6709\u4ec0\u4e48\u529e\u6cd5\u53ef\u4ee5\u5728\u4f4e\u4ee3\u4ef7\u4e0b\u4ecd\u7136\u5f97\u5230\u7cbe\u7ec6\u7684\u5206\u5272\u7ed3\u679c\u5462\uff1f\u5176\u5b9e\u5f88\u91cd\u8981\u7684\u4e00\u70b9\u662f\u5f80\u5f80\u8fd9\u4e9b\u4e0d\u51c6\u786e\u7684\u90e8\u5206\u662f\u5728\u7269\u4f53\u7684\u8fb9\u7f18\uff0c\u8fd9\u4e9b\u8fb9\u7f18\u5176\u5b9e\u53ea\u5360\u4e86\u6574\u4e2a\u7269\u4f53\u4e2d\u975e\u5e38\u5c0f\u7684\u4e00\u90e8\u5206\u3002\u6240\u4ee5\u57fa\u4e8e\u8fd9\u6837\u7684\u4e00\u4e2a\u60f3\u6cd5\uff0c\u4f5c\u8005\u63d0\u51fa\u53ef\u4ee5\u6bcf\u6b21\u5728\u9884\u6d4b\u51fa\u6765\u7684mask\u4e2d\u53ea\u9009\u62e9Top N\u6700\u4e0d\u786e\u5b9a\u7684\u4f4d\u7f6e\u8fdb\u884c\u7ec6\u5206\u9884\u6d4b\u3002<\/strong>\u6bcf\u4e2a\u7ec6\u5206\u70b9\u7684\u7279\u5f81\u53ef\u4ee5\u901a\u8fc7Bilinear\u63d2\u503c\u5f97\u5230\uff0c\u6bcf\u4e2a\u4f4d\u7f6e\u4e0a\u7684classifier\u901a\u8fc7\u4e00\u4e2a\u7b80\u5355\u7684MLP\u6765\u5b9e\u73b0\u3002\u8fd9\u5176\u5b9e\u662f\u7b49\u4ef7\u4e8e\u7528\u4e00\u4e2a1*1\u7684conv\u6765\u9884\u6d4b\uff0c\u4f46\u662f\u5bf9\u4e8e\u4e2d\u5fc3\u5f88\u786e\u5b9a\u7684\u70b9\u5e76\u4e0d\u8ba1\u7b97\u3002\u6574\u4f53\u7684\u793a\u610f\u56fe\u5982\u4e0b\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

PointRend \u89e3\u51b3\u4e86\u4ec0\u4e48\u95ee\u9898\uff1f<\/strong><\/h2>\n\n\n\n

\u8fd9\u7bc7\u8bba\u6587\u8bb2\u4e86\u4e00\u4e2a\u5f88\u597d\u542c\u7684\u6545\u4e8b\uff0c\u5373\uff1a\u628a\u8bed\u4e49\u5206\u5272\u4ee5\u53ca\u5b9e\u4f8b\u5206\u5272\u95ee\u9898\uff08\u7edf\u79f0\u56fe\u50cf\u5206\u5272\u95ee\u9898\uff09\u5f53\u505a\u4e00\u4e2a\u6e32\u67d3\u95ee\u9898\u6765\u89e3\u51b3\u3002\u6545\u4e8b\u867d\u7136\u8fd9\u4e48\u8bb2\uff0c\u4f46\u672c\u8d28\u4e0a\u8fd9\u7bc7\u8bba\u6587\u5176\u5b9e\u662f\u4e00\u4e2a\u65b0\u578b\u4e0a\u91c7\u6837\u65b9\u6cd5\uff0c\u9488\u5bf9\u7269\u4f53\u8fb9\u7f18\u7684\u56fe\u50cf\u5206\u5272\u8fdb\u884c\u4f18\u5316\uff0c\u4f7f\u5176\u5728\u96be\u4ee5\u5206\u5272\u7684\u7269\u4f53\u8fb9\u7f18\u90e8\u5206\u6709\u66f4\u597d\u7684\u8868\u73b0<\/strong>\u3002<\/p>\n\n\n\n

\u4f5c\u4e3a\u4e00\u4e2a\u5c0f\u767d\uff0c\u90a3\u4e48\u95ee\u9898\u6765\u4e86\uff1a<\/p>\n\n\n\n

1\u3001\u4ec0\u4e48\u662f\u6e32\u67d3\uff1f<\/p>\n\n\n\n

2\u3001\u4e3a\u4ec0\u4e48\u8981\u628a\u56fe\u50cf\u5206\u5272\u95ee\u9898\u5f53\u505a\u6e32\u67d3\u95ee\u9898\u5462\uff1f<\/p>\n\n\n\n

\u8981\u60f3\u77e5\u9053\u4ec0\u4e48\u662f\u6e32\u67d3\uff0c\u53ef\u4ee5\u53c2\u8003\uff1a<\/p>\n\n\n\n

\u8ba1\u7b97\u673a\u4e2d\u6240\u8bf4\u7684\u300c\u6e32\u67d3\u300d\u662f\u4ec0\u4e48\u610f\u601d\uff1f<\/a><\/p>\n\n\n\n

\u7b80\u5355\u6765\u8bf4\uff0c\u6e32\u67d3\u5c31\u662f\u201c\u7ed8\u5236\u201d\uff0c\u628a3D\u7684\u7269\u4f53\u57282D\u5e73\u9762\u4e0a\u7ed8\u5236\u51fa\u6765\u3002<\/p>\n\n\n\n

\u4e3a\u4ec0\u4e48\u8981\u628a\u56fe\u50cf\u5206\u5272\u95ee\u9898\u548c\u6e32\u67d3\u95ee\u9898\u626f\u5728\u4e00\u8d77\u5462\uff1f\u56e0\u4e3a\u8bb2\u6545\u4e8b\u597d\u542c\u554a\uff0c\u8bba\u6587\u597d\u5199\u561b….\u54b3\u54b3…\u4e0d\u4e0d\uff0c\u662f\u56e0\u4e3a\u4e8c\u8005\u6709\u7c7b\u4f3c\u7684\u95ee\u9898\u8981\u89e3\u51b3\uff1a\u5373\u7269\u4f53\u8fb9\u7f18\u96be\u4ee5\u5904\u7406\u3002<\/p>\n\n\n\n

\u5177\u4f53\u6765\u8bf4\uff0c\u5728\u56fe\u50cf\u6e32\u67d3\u4e2d\uff0c\u5bf9\u4e8e\u591a\u4e2a3D\u7269\u4f53\uff0c\u5728\u8fb9\u7f18\u8981\u5224\u65ad\u5bf9\u4e8e\u955c\u5934\u800c\u8a00\u8c01\u5148\u8c01\u540e\uff0c\u800c\u4e14\u8fd8\u5f97\u6297\u952f\u9f7f\uff1b\u800c\u5bf9\u4e8e\u56fe\u50cf\u5206\u5272\u95ee\u9898\uff0c\u8fb9\u7f18\u6062\u590d\u4e5f\u4e00\u76f4\u662f\u4e2a\u9ebb\u70e6\u4e8b\u513f\uff0c\u56e0\u4e3a\u5728\u5178\u578b\u7684\u8bed\u4e49\u5206\u5272\u7f51\u7edc\u4e2d\uff08\u5982FCN\u3001DeepLab\uff09\uff0c\u5728CNN\u5185\u90e8\u4e00\u822c\u90fd\u4f1a\u76f8\u5bf9\u8f93\u5165\u56fe\u50cf\u964d\u91c7\u683716\u500d\uff0c\u7136\u540e\u518d\u60f3\u529e\u6cd5\u4e0a\u91c7\u6837\u56de\u53bb\u3002\u66f4\u7ec6\u81f4\u5730\u8bf4\uff0c\u5bf9\u4e8e DeepLabV3+\uff0c\u6a21\u578b\u6700\u540e\u76f4\u63a5\u662f\u4e00\u4e2a4\u500d\u7684\u53cc\u7ebf\u6027\u63d2\u503c\u4e0a\u91c7\u6837\uff0c\u8fd9\u663e\u7136\u5bf9\u7269\u4f53\u8fb9\u7f18\u7684\u9884\u6d4b\u5341\u5206\u4e0d\u5229\u3002\u867d\u7136 DeepLabV3+\u5f53\u65f6\u57282017\u5e74\u5c31\u8fbe\u5230\u4e86\u79d2\u5929\u79d2\u5730\u7684 89%mIoU on VOC2012 test \uff08\u4f7f\u7528\u4e86300M JFT \u6570\u636e\u96c6\u9884\u8bad\u7ec3\uff09\uff0c\u81f3\u4eca\u65e0\u4eba\u8d85\u8d8a\uff08\u56e0\u4e3aJFT \u6570\u636e\u96c6 Google\u6ca1\u6709\u516c\u5f00 \\\u624b\u52a8\u6ed1\u7a3d\uff09\uff0c\u4f46\u663e\u7136\u8fd9\u4e2a\u4e0a\u91c7\u6837\u8fc7\u7a0b\u4ecd\u7136\u5b58\u5728\u8f83\u5927\u7684\u63d0\u5347\u7a7a\u95f4\u3002<\/p>\n\n\n\n

\u53c2\u8003\u94fe\u63a5\uff1aUno Whoiam\uff1aDeepLab \u8bed\u4e49\u5206\u5272\u6a21\u578b v1\u3001v2\u3001v3\u3001v3+ \u6982\u8981\uff08\u9644 Pytorch \u5b9e\u73b0\uff09<\/a><\/p><\/blockquote>\n\n\n\n

\u800c\u5728\u5b9e\u4f8b\u5206\u5272\u7f51\u7edc\u4e2d\uff0cMask R-CNN \u8fd9\u8d27\u751f\u6210\u7684 Mask \u624d 28×28\uff0c\u8981\u662f\u628a\u8fd9\u6837\u7684 mask \u62c9\u4f38\u5230 \u4e0d\u8bf4\u591a\u4e86\u6bd4\u5982 256×256\uff0c\u8fd8\u6307\u671b\u5b83\u53ef\u4ee5\u5f88\u597d\u5730\u9884\u6d4b\u8fb9\u7f18\uff1f\u6211\u53ea\u80fd\u8bf4\u8fd9\u662f\u5728\u60f3Peach\u3002<\/p>\n\n\n\n

\u4e8b\u5b9e\u4e0a\uff0c\u5728\u56fe\u50cf\u5206\u5272\u4efb\u52a1\u4e0a\u8fb9\u7f18\u9884\u6d4b\u4e0d\u7406\u60f3\u8fd9\u4e2a\u60c5\u51b5\u5176\u5b9e\u5728\u8bb8\u591a\u524d\u4eba\u7684\u5de5\u4f5c\u4e2d\u90fd\u6709\u63d0\u53ca\uff0c\u6bd4\u5982 Not All Pixels Are Equal: Difficulty-Aware Semantic Segmentation via Deep Layer Cascade \u4e2d\u5c31\u8be6\u7ec6\u7edf\u8ba1\u4e86\u8bed\u4e49\u5206\u5272\u4e2d\uff0c\u6a21\u578b\u6700\u5bb9\u6613\u8bef\u5224\u7684 pixel\u57fa\u672c\u4e0a\u90fd\u5728\u7269\u4f53\u8fb9\u7f18\uff08\u5982\u4e0b\u56fe\u53f3\u4e0a\u7ea2\u8272\u90e8\u5206\u6807\u8bb0\uff09 \u3002<\/p>\n\n\n\n

\"\"\/<\/figure>\n\n\n\n

\u800c\u5173\u4e8e\u4e0a\u91c7\u6837\u5176\u5b9e\u4e5f\u6709\u4e00\u4e9b\u524d\u4eba\u7684\u5de5\u4f5c\uff0c\u5982 Decoders Matter for Semantic Segmentation: Data-Dependent Decoding Enables Flexible Feature Aggregation\uff0c\u5728\u5b9e\u73b0\u4e0a\u6709\u70b9\u50cf\u8d85\u5206\u8fa8\u7387\u7f51\u7edc ESPCN \u91cc\u4f7f\u7528\u7684 sub-pixel convolutional layer \u7684\u64cd\u4f5c\uff0c\u4e0d\u8fc7\u591a\u52a0\u4e86\u4e00\u4e2a\u4e8c\u9636\u8303\u6570\u7ea6\u675f\uff1a<\/p>\n\n\n\n

\"\"\/<\/figure>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n

\u603b\u7684\u6765\u8bf4\uff0c\u56fe\u50cf\u5206\u5272\u8fb9\u7f18\u9884\u6d4b\u662f\u4e00\u4e2a\u672a\u88ab\u5f88\u597d\u89e3\u51b3\u7684\u95ee\u9898\uff0c\u800c\u4f55\u607a\u660e\u56e2\u961f\u7684 PointRend \u662f\u5bf9\u6b64\u95ee\u9898\u7684\u4e00\u4e2a\u65b0\u7684\u601d\u8def\u548c\u89e3\u6cd5\uff0c\u63a5\u4e0b\u6765\u5c06\u4ecb\u7ecd PointRend \u662f\u5982\u4f55 work \u7684\u3002<\/p>\n\n\n\n

\u6587\u4e3b\u8981\u8d21\u732e<\/strong><\/h2>\n\n\n\n

1.\u63d0\u51fa\u53ef\u5d4c\u5165\u4e3b\u6d41\u7f51\u7edc\u7684PointRend\u6a21\u5757\uff0c\u63d0\u9ad8\u4e86\u56fe\u50cf\u5206\u5272\u7cbe\u5ea6\u3002<\/p>\n\n\n\n

2.\u628a\u56fe\u50cf\u5206\u5272\u95ee\u9898\u770b\u4f5c\u6e32\u67d3\u95ee\u9898\uff0c\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u65b0\u578b\u4e0a\u91c7\u6837\u65b9\u6cd5\uff0c\u4e3a\u56fe\u50cf\u5206\u5272\u63d0\u4f9b\u72ec\u7279\u89c6\u89d2\u3002<\/p>\n\n\n\n

3.\u964d\u4f4e\u4e86\u8bad\u7ec3\u6240\u9700\u7684\u7b97\u529b\u3002\u8f93\u51fa224\u00d7224\u5206\u8fa8\u7387\u56fe\u50cf\uff0cPointRend\u53ea\u97000.9B FLOPs\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u603b\u4f53\u601d\u8def<\/strong><\/h2>\n\n\n\n

PointRend \u65b9\u6cd5\u8981\u70b9\u603b\u7ed3\u6765\u8bf4\u662f\u4e00\u4e2a\u8fed\u4ee3\u4e0a\u91c7\u6837\u7684\u8fc7\u7a0b\uff1a<\/p>\n\n\n\n

while \u8f93\u51fa\u7684\u5206\u8fa8\u7387 < \u56fe\u7247\u5206\u8fa8\u7387\uff1a<\/p>\n\n\n\n

  1. \u5bf9\u8f93\u51fa\u7ed3\u679c\u8fdb\u884c2\u500d\u53cc\u7ebf\u6027\u63d2\u503c\u4e0a\u91c7\u6837\u5f97\u5230 coarse prediction_i\u3002\uff08\u7c97\u5206\u8fa8\u7387\u9884\u6d4b\uff09<\/strong><\/li>
  2. \u6311\u9009\u51fa N \u4e2a\u201c\u96be\u70b9\u201d\uff0c\u5373\u7ed3\u679c\u5f88\u6709\u53ef\u80fd\u548c\u5468\u56f4\u70b9\u4e0d\u4e00\u6837\u7684\u70b9\uff08\u4f8b\u5982\u7269\u4f53\u8fb9\u7f18\uff09\u3002<\/strong><\/li>
  3. \u5bf9\u4e8e\u6bcf\u4e2a\u96be\u70b9\uff0c\u83b7\u53d6\u5176\u201c\u8868\u5f81\u5411\u91cf\u201d\uff0c\u201c\u8868\u5f81\u5411\u91cf\u201d\u7531\u4e24\u4e2a\u90e8\u5206\u7ec4\u6210\uff0c\u5176\u4e00\u662f\u4f4e\u5c42\u7279\u5f81\uff08fine-grained features\uff09\uff0c\u901a\u8fc7\u4f7f\u7528\u70b9\u7684\u5750\u6807\uff0c\u5728\u4f4e\u5c42\u7684\u7279\u5f81\u56fe\u4e0a\u8fdb\u884c\u53cc\u7ebf\u6027\u63d2\u503c\u83b7\u5f97\uff08\u7c7b\u4f3c RoI Align\uff09\uff0c\u5176\u4e8c\u662f\u9ad8\u5c42\u7279\u5f81\uff08coarse prediction\uff09\uff0c\u7531\u6b65\u9aa4 1 \u83b7\u5f97\u3002<\/strong><\/li>
  4. \u4f7f\u7528 MLP \u5bf9\u201c\u8868\u5f81\u5411\u91cf\u201d\u8ba1\u7b97\u5f97\u5230\u65b0\u7684\u9884\u6d4b\uff0c\u66f4\u65b0 coarse prediction_i \u5f97\u5230 coarse prediction_i+1\u3002\u8fd9\u4e2a MLP \u5176\u5b9e\u53ef\u4ee5\u770b\u505a\u4e00\u4e2a\u53ea\u5bf9\u201c\u96be\u70b9\u201d\u7684\u201c\u8868\u5f81\u5411\u91cf\u201d\u8fdb\u884c\u8fd0\u7b97\u7684\u7531\u591a\u4e2a conv1x1 \u7ec4\u6210\u7684\u5c0f\u7f51\u7edc\u3002<\/strong><\/li><\/ol>\n\n\n\n

    \u6574\u4e2a\u8fc7\u7a0b\u53ef\u4ee5\u8fd9\u4e48\u7406\u89e3\uff1a<\/p>\n\n\n\n

    \u5c0f\u660e\u540c\u5b66\u505a\u9898\uff0c\u73b0\u5728\u6709\u5df2\u77e5\u6761\u4ef6\uff08coarse prediction_0\uff0cfine-grained features\uff09,\u60f3\u6c42\u89e3\u7b54\u6848\uff08coarse prediction_k\uff09\uff0c\u53d1\u73b0\u76f4\u63a5\u6c42\uff08\u53cc\u7ebf\u6027\u63d2\u503cor\u5176\u5b83\u65b9\u6cd5\uff09\u4e0d\u591f\u51c6\u786e\uff0c\u90a3\u5c31\u4e00\u6b65\u4e00\u6b65\u6765\u5427\uff08\u4ececoarse prediction_1\uff0ccoarse prediction_2….\u6c42\u5230coarse prediction_k\uff09\u3002\u597d\u7684\uff0c\u73b0\u5728\u6c42coarse prediction_1\uff0c\u8bf6\uff0c\u53d1\u73b0\u6709\u597d\u591a\u4e1c\u897f\u4e0d\u77e5\u9053\uff0c\u4e0d\u80fd\u4ece coarse prediction_0 \u76f4\u63a5\u5f97\u5230\u600e\u4e48\u529e\uff1f\u90a3\u5c31\u627e\u51fa\u4e0d\u77e5\u9053\u7684\uff08\u201c\u96be\u70b9\u201d\uff09\uff0c\u5728 fine-grained features \u91cc\u9762\u627e\u51fa\u5bf9\u5e94\u7684\u7ebf\u7d22\uff08ROIAlign-like \u53cc\u7ebf\u6027\u63d2\u503c\uff09\uff0c\u7136\u540e\u5728\u7ed3\u5408 coarse prediction_0 \u5f97\u5230\u6574\u4f53\u7ebf\u7d22\uff08\u201c\u7279\u5f81\u5411\u91cf\u201d\uff09\u6c42\u89e3\uff08\u4f7f\u7528MLP\u8ba1\u7b97\uff09\uff0c\u55ef\uff0c\u7ec8\u4e8e\u5f97\u5230 coarse prediction_1\u4e86\u3002\u518d\u7528\u540c\u6837\u7684\u601d\u8def\u53cd\u590d\u6c42\u89e3\uff0c\u76f4\u5230 coarse prediction_k\u3002<\/p>\n\n\n\n

    \u793a\u610f\u56fe\u5982\u4e0b\uff1a<\/p>\n\n\n\n

    \"\"\/<\/figure><\/div>\n\n\n\n
    \"\"\/
    \u5bf9\u4e8e\u4e00\u4e2acoarse prediction(4×4\u5927\u5c0f)\uff0c\u5c06\u5176\u4e0a\u91c7\u6837\u4e24\u500d(8×8\u5927\u5c0f\uff0c\u8fd9\u91cc\u53ef\u4ee5\u7406\u89e3\u4e3a\u68c0\u6d4b\u5934\u7684\u8f93\u51fa)\u540e\uff0c\u53d6\u4e86\u4e00\u4e9b\u96be\u5206\u5272\u7684\u70b9\uff08\u5927\u591a\u662f\u8fb9\u7f18\u90e8\u5206\uff09\uff0c\u53d6\u8fd9\u4e9b\u70b9\u7684\u7279\u5f81\u5411\u91cf\u8f93\u5165\u5230MLP\u7f51\u7edc\u4e2d\uff0c\u8fdb\u884cpoint prediction\uff0c\u5f97\u5230\u6bcf\u4e00\u4e2a\u70b9\u7684\u65b0\u7c7b\u522b\uff0c\u6700\u540e\u7ed3\u679c\u8f93\u51fa(8×8\u5927\u5c0f\uff0c\u8fb9\u7f18\u66f4\u52a0\u7cbe\u786e\u7684\u7ed3\u679c)\u3002<\/figcaption><\/figure>\n\n\n\n

    \u53e6\u5916\uff0c\u5176PointRend \u8bad\u7ec3\u4e3a\u4e86\u8282\u7701\u65f6\u95f4\uff0c\u6ca1\u6709\u4f7f\u7528\u4e0a\u8ff0\u7684\u8fed\u4ee3\u8fc7\u7a0b\uff0c\u800c\u662f\u4f7f\u7528\u591a\u79cd\u7ec4\u5408\u7684\u91c7\u6837\u65b9\u6cd5\uff0c\u4e0d\u8d58\u8ff0\uff0c\u8be6\u89c1paper\u3002<\/p>\n\n\n\n

    1. \u4ecePointRend\u7684\u5e94\u7528\u601d\u8def\u4e2d\u53ef\u4ee5\u770b\u5230\uff0c\u8fd9\u91cc\u5305\u542b\u4e86\u4e24\u4e2a\u9636\u6bb5\u7684\u7279\u5f81\u5904\u7406\uff0c\u5206\u522b\u662ffine-grained features\u548ccoarse prediction\u90e8\u5206\uff0c\u5982\u679c\u4e3b\u5e72\u7f51\u7edc\u662fResNet\uff0c\u90a3\u4e48fine-grained features\u5c31\u662fResNet\u7684stage2\u8f93\u51fa\uff0c\u4e5f\u5c31\u662f4\u500d\u4e0b\u91c7\u6837\u65f6\u7684\u7cbe\u7ec6\u5206\u5272\u7ed3\u679c\uff0c\u800ccoarse prediction\u5c31\u662f\u68c0\u6d4b\u5934\u7684\u9884\u6d4b\u7ed3\u679c\uff08\u8fd8\u672a\u4e0a\u91c7\u6837\u8fd8\u539f\u6210\u539f\u56fe\u7684\u7ed3\u679c\uff09\u3002<\/li>
    2. \u4ececoarse prediction\u4e2d\u6311\u9009N\u4e2a\u201c\u96be\u70b9\u201d\uff0c\u4e5f\u5c31\u662f\u7ed3\u679c\u5f88\u6709\u53ef\u80fd\u548c\u5468\u56f4\u70b9\u4e0d\u4e00\u6837\u7684\u70b9\uff08\u6bd4\u5982\u7269\u4f53\u8fb9\u7f18\u7684\u70b9\uff09\u3002\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u96be\u70b9\uff0c\u83b7\u53d6\u4ed6\u7684\u201c\u7279\u5f81\u5411\u91cf\u201d\uff0c\u5bf9\u4e8e\u70b9\u7279\u5f81\u5411\u91cf\uff08point features\uff09\uff0c\u4e3b\u8981\u7531\u4e24\u90e8\u5206\u7ec4\u6210\uff0c\u5206\u522b\u662ffine-grained features\u7684\u5bf9\u5e94\u70b9\u548ccoarse prediction\u7684\u5bf9\u5e94\u70b9\u7684\u7279\u5f81\u5411\u91cf\uff0c\u5c06\u8fd9\u4e2a\u4e24\u4e2a\u7279\u5f81\u5411\u91cf\u62fc\u63a5\u6210\u4e00\u4e2a\u5411\u91cf\u3002<\/li>
    3. \u63a5\u7740\uff0c\u901a\u8fc7\u4e00\u4e2aMLP\u7f51\u7edc\u5bf9\u8fd9\u4e2a\u201c\u7279\u5f81\u5411\u91cf\u201d\u8fdb\u884c\u9884\u6d4b\uff0c\u66f4\u65b0coarse prediction\u3002\u4e5f\u5c31\u76f8\u5f53\u4e8e\u5bf9\u8fd9\u4e2a\u96be\u70b9\u8fdb\u884c\u65b0\u7684\u9884\u6d4b\uff0c\u5bf9\u4ed6\u8fdb\u884c\u5206\u7c7b\u3002<\/li><\/ol>\n\n\n\n

      \u770b\u5b8c\u8fd9\u4e2a\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u8fd9\u4e48\u7406\u89e3\uff0c\u5c06\u9884\u6d4b\u96be\u7684\u70b9\uff08\u8fb9\u7f18\u70b9\uff09\u63d0\u53d6\u51fa\u6765\uff0c\u518d\u63d0\u53d6\u5176\u7279\u5f81\u5411\u91cf\uff0c\u7ecf\u8fc7MLP\u7f51\u7edc\uff0c\u5c06\u8fd9\u4e2a\u70b9\u7684\u5f52\u5c5e\u8fdb\u884c\u5206\u7c7b\uff0c\u7136\u540e\u63d0\u5347\u8fd9\u4e9b\u70b9\u7684\u5206\u7c7b\u51c6\u786e\u7387\u3002\u8fd9\u5c31\u662fPointRend\u7684\u601d\u60f3\u3002<\/p>\n\n\n\n

      \u4e00\u4e2aPointRend\u6a21\u5757\u5305\u62ec\u4e09\u90e8\u5206<\/strong>\u3002<\/p>\n\n\n\n

      1.point selection strategy\uff1a\u7528\u4e8einference\u548ctraing\u7684\u70b9\u9009\u62e9<\/u><\/strong><\/p>\n\n\n\n

      \u5bf9\u4e8e\u70b9\u91c7\u6837\u8fc7\u7a0b\uff0c\u9700\u8981\u5bf9\u6a21\u578b\u7684Train\u8fc7\u7a0b\u548cInference\u8fc7\u7a0b\u505a\u533a\u5206<\/p>\n\n\n\n

      \u8be5\u65b9\u6cd5\u7684\u6838\u5fc3\u601d\u60f3\u662f\u7075\u6d3b\u81ea\u9002\u5e94\u5730\u9009\u62e9\u56fe\u50cf\u5e73\u9762\u4e0a\u7684\u70b9\u6765\u9884\u6d4b\u5206\u5272\u6807\u7b7e<\/strong>\u3002\u76f4\u89c2\u5730\u8bf4\uff0c\u8fd9\u4e9b\u70b9\u5e94\u8be5\u66f4\u5bc6\u96c6\u5730\u4f4d\u4e8e\u9ad8\u9891\u533a\u57df\u9644\u8fd1\uff0c\u4f8b\u5982\u7269\u4f53\u8fb9\u754c\uff0c\u7c7b\u4f3c\u4e8e\u5c04\u7ebf\u8ffd\u8e2a\u4e2d\u7684\u53cd\u6df7\u53e0\u95ee\u9898\u3002\u6211\u4eec\u4ea7\u751f\u4e86\u63a8\u7406<\/strong>\u548c\u8bad\u7ec3<\/strong>\u7684\u60f3\u6cd5\u3002<\/p>\n\n\n\n