{"id":9021,"date":"2022-10-21T19:22:38","date_gmt":"2022-10-21T11:22:38","guid":{"rendered":"http:\/\/139.9.1.231\/?p=9021"},"modified":"2022-10-21T20:11:36","modified_gmt":"2022-10-21T12:11:36","slug":"pointnet","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/10\/21\/pointnet\/","title":{"rendered":"pointnet–\u57fa\u4e8e\u70b9\u4e91\u7684\u5206\u7c7b\u548c\u5206\u5272\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5"},"content":{"rendered":"\n

\u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/1612.00593<\/a>\uff08cvpr2017\uff09<\/p>\n\n\n\n

code\uff1ahttps:\/\/github.com\/charlesq34\/pointnet<\/a><\/p>\n\n\n\n

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

\u57fa\u7840\u77e5\u8bc6\uff1a<\/h3>\n\n\n\n

1\u3001\u4ec0\u4e48\u662f\u70b9\u4e91\uff1f<\/p>\n\n\n\n

\u7b80\u5355\u6765\u8bf4\u5c31\u662f\u4e00\u5806\u4e09\u7ef4\u70b9\u7684\u96c6\u5408\uff0c\u5fc5\u987b\u5305\u62ec\u5404\u4e2a\u70b9\u7684\u4e09\u7ef4\u5750\u6807\u4fe1\u606f\uff0c\u5176\u4ed6\u4fe1\u606f\u6bd4\u5982\u5404\u4e2a\u70b9\u7684\u6cd5\u5411\u91cf\u3001\u989c\u8272\u7b49\u5747\u662f\u53ef\u9009<\/strong>\u3002\u70b9\u4e91\u7684\u6587\u4ef6\u683c\u5f0f\u53ef\u4ee5\u6709\u5f88\u591a\u79cd\uff0c\u5305\u62ecxyz\uff0cnpy\uff0cply\uff0cobj\uff0coff\u7b49\uff08\u6709\u4e9b\u662fmesh\u4e0d\u8fc7\u95ee\u9898\u4e0d\u5927\uff0c\u56e0\u4e3amesh\u53ef\u4ee5\u901a\u8fc7\u6cca\u677e\u91c7\u6837\u7b49\u65b9\u5f0f\u8f6c\u5316\u6210\u70b9\u4e91\uff09\u3002\u5bf9\u4e8e\u5355\u4e2a\u70b9\u4e91\uff0c\u5982\u679c\u4f60\u4f7f\u7528np.loadtxt\u5f97\u5230\u7684\u5b9e\u9645\u4e0a\u5c31\u662f\u4e00\u4e2a\u7ef4\u5ea6\u4e3a (num_points,num_channels) \u7684\u5f20\u91cf\uff0cnum_channels\u4e00\u822c\u4e3a3\uff0c\u8868\u793a\u70b9\u4e91\u7684\u4e09\u7ef4\u5750\u6807\u3002<\/p>\n\n\n\n

\u8fd9\u91cc\u4ee5horse.xyz\u6587\u4ef6\u4e3a\u4f8b\uff0c\u5b9e\u9645\u5c31\u662f\u6587\u672c\u6587\u4ef6\uff0c\u6253\u5f00\u540e\u6570\u636e\u957f\u8fd9\u6837\uff08\u5c40\u90e8\uff0c\u603b\u5171\u67092048\u4e2a\u70b9\uff09\uff1a<\/p>\n\n\n\n

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

\u5b9e\u9645\u5c31\u662f\u4e00\u5806\u70b9\u7684\u4fe1\u606f\uff0c\u8fd9\u91cc\u53ea\u6709\u4e09\u7ef4\u5750\u6807\uff0c\u5c06\u5176\u53ef\u89c6\u5316\u51fa\u6765\u957f\u8fd9\u6837\uff1a<\/p>\n\n\n\n

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

2\u3001\u70b9\u4e91\u5904\u7406\u4efb\u52a1\u662f\u91cd\u8981\u7684<\/strong>\uff1f<\/p>\n\n\n\n

\u4e09\u7ef4\u56fe\u5f62\u5177\u6709\u591a\u79cd\u8868\u73b0\u5f62\u5f0f\uff0c\u5305\u62ec\u4e86mesh\u3001\u4f53\u7d20\u3001\u70b9\u4e91\u7b49\uff0c\u751a\u81f3\u8fd8\u6709\u4e9b\u65b9\u6cd5\u4f7f\u7528\u591a\u89c6\u56fe\u6765\u5bf9\u4e09\u7ef4\u56fe\u5f62\u8868\u5f81\u3002\u800c\u70b9\u4e91\u5728\u4ee5\u4e0a\u5404\u79cd\u5f62\u5f0f\u7684\u6570\u636e\u4e2d\u7b97\u662f\u65e5\u5e38\u751f\u6d3b\u4e2d\u6700\u80fd\u591f\u5927\u89c4\u6a21\u83b7\u53d6\u548c\u4f7f\u7528\u7684\u6570\u636e\u7ed3\u6784\u4e86\uff0c\u5305\u62ec\u81ea\u52a8\u9a7e\u9a76\u3001\u589e\u5f3a\u73b0\u5b9e\u7b49\u5728\u5185\u7684\u5e94\u7528\u9700\u8981\u76f4\u63a5\u6216\u95f4\u63a5\u4ece\u70b9\u4e91\u4e2d\u63d0\u53d6\u4fe1\u606f\uff0c\u70b9\u4e91\u5904\u7406\u4e5f\u9010\u6e10\u6210\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u975e\u5e38\u91cd\u8981\u7684\u4e00\u90e8\u5206\u3002<\/p>\n\n\n\n

\u6b63\u6587\uff1a<\/h3>\n\n\n\n

PointNet\u6240\u4f5c\u7684\u4e8b\u60c5\u5c31\u662f\u5bf9\u70b9\u4e91\u505a\u7279\u5f81\u5b66\u4e60\uff0c\u5e76\u5c06\u5b66\u4e60\u5230\u7684\u7279\u5f81\u53bb\u505a\u4e0d\u540c\u7684\u5e94\u7528\uff1a\u5206\u7c7b\uff08shape-wise feature\uff09\u3001\u5206\u5272\uff08point-wise feature\uff09\u7b49\u3002<\/p>\n\n\n\n

PointNet\u4e4b\u6240\u4ee5\u5f71\u54cd\u529b\u5de8\u5927\uff0c\u5c31\u662f\u56e0\u4e3a\u5b83\u4e3a\u70b9\u4e91\u5904\u7406\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7b80\u5355\u3001\u9ad8\u6548\u3001\u5f3a\u5927\u7684\u7279\u5f81\u63d0\u53d6\u5668\uff08encoder\uff09\uff0c\u51e0\u4e4e\u53ef\u4ee5\u5e94\u7528\u5230\u70b9\u4e91\u5904\u7406\u7684\u5404\u4e2a\u5e94\u7528\u4e2d\uff0c\u5176\u5730\u4f4d\u7c7b\u4f3c\u4e8e\u56fe\u50cf\u9886\u57df\u7684AlexNet\u3002<\/p>\n\n\n\n

1\u3001\u52a8\u673a<\/h4>\n\n\n\n

\u70b9\u4e91\u6216\u8005mesh\uff0c\u5927\u591a\u6570\u7814\u7a76\u4eba\u5458\u90fd\u662f\u5c06\u5176\u8f6c\u5316\u62103D\u4f53\u7d20\u6216\u8005\u591a\u89c6\u56fe\u6765\u505a\u7279\u5f81\u5b66\u4e60\u7684\uff0c\u8fd9\u5176\u4e2d\u7684\u5de5\u4f5c\u5305\u62ec\u4e86VoxelNet, MVCNN\u7b49\u3002\u8fd9\u4e9b\u5de5\u4f5c\u90fd\u6216\u591a\u6216\u5c11\u5b58\u5728\u4e86\u4e00\u4e9b\u95ee\u9898\u3002<\/p>\n\n\n\n

\u76f4\u63a5\u5bf9\u70b9\u4e91\u505a\u7279\u5f81\u5b66\u4e60\u4e5f\u4e0d\u662f\u4e0d\u53ef\u4ee5\uff0c\u4f46\u6709\u51e0\u4e2a\u95ee\u9898\u9700\u8981\u8003\u8651\uff1a\u7279\u5f81\u5b66\u4e60\u9700\u8981\u5bf9\u70b9\u4e91\u4e2d\u5404\u4e2a\u70b9\u7684\u6392\u5217\u4fdd\u6301\u4e0d\u53d8\u6027\u3001\u7279\u5f81\u5b66\u4e60\u9700\u8981\u5bf9rigid transformation\u4fdd\u6301\u4e0d\u53d8\u6027\u7b49\u3002\u867d\u7136\u6709\u6311\u6218\uff0c\u4f46\u662f\u6df1\u5ea6\u5b66\u4e60\u5f3a\u5927\u7684\u8868\u5f81\u80fd\u529b\u4ee5\u53ca\u5176\u5728\u56fe\u50cf\u9886\u57df\u53d6\u5f97\u7684\u5de8\u5927\u6210\u529f\uff0c\u56e0\u6b64\u662f\u5f88\u6709\u5fc5\u8981\u76f4\u63a5\u5728\u70b9\u4e91\u4e0a\u8fdb\u884c\u5c1d\u8bd5\u7684\u3002<\/p>\n\n\n\n

2\u3001\u8d21\u732e<\/h4>\n\n\n\n
  1. \u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u65b0\u9896\u7684\u6df1\u5c42\u7f51\u7edc\u67b6\u6784\u6765\u5904\u7406\u4e09\u7ef4\u4e2d\u7684\u65e0\u5e8f\u70b9\u96c6<\/li>
  2. \u6211\u4eec\u8bbe\u8ba1\u7684\u7f51\u7edc\u8868\u5f81\u53ef\u4ee5\u505a\u4e09\u7ef4\u56fe\u5f62\u5206\u7c7b\u3001\u56fe\u5f62\u7684\u5c40\u90e8\u5206\u5272\u4ee5\u53ca\u573a\u666f\u7684\u8bed\u4e49\u5206\u5272\u7b49\u4efb\u52a1<\/li>
  3. \u6211\u4eec\u63d0\u4f9b\u4e86\u5b8c\u5907\u7684\u7ecf\u9a8c\u548c\u7406\u8bba\u5206\u6790\u6765\u8bc1\u660ePointNet\u7684\u7a33\u5b9a\u548c\u9ad8\u6548\u3002<\/li>
  4. \u5145\u5206\u7684\u6d88\u878d\u5b9e\u9a8c\uff0c\u8bc1\u660e\u7f51\u7edc\u5404\u4e2a\u90e8\u5206\u5bf9\u4e8e\u8868\u5f81\u7684\u6709\u6548\u6027\u3002<\/li><\/ol>\n\n\n\n

    3\u3001\u65b9\u6cd5<\/h4>\n\n\n\n

    3.1 \u70b9\u4e91\u7684\u51e0\u4e2a\u7279\u70b9\uff1a<\/p>\n\n\n\n

    1. \u65e0\u5e8f\u6027 –> \u5bf9\u79f0\u51fd\u6570\u8bbe\u8ba1\u7528\u4e8e\u8868\u5f81<\/strong><\/li>
    2. \u70b9\u4e0d\u662f\u5b64\u7acb\u7684\uff0c\u9700\u8981\u8003\u8651\u5c40\u90e8\u7ed3\u6784 –> \u5c40\u90e8\u5168\u5c40\u7279\u5f81\u7ed3\u5408<\/strong><\/li>
    3. \u4eff\u5c04\u53d8\u6362\u65e0\u5173\u6027 –> alignment network<\/strong><\/li><\/ol>\n\n\n\n

      \uff08\u91cd\u8981\uff09\u5173\u4e8e\u7b2c\u4e09\u70b9<\/strong>\uff1a\u76f8\u540c\u7684\u70b9\u4e91\u5728\u7a7a\u95f4\u4e2d\u7ecf\u8fc7\u4e00\u5b9a\u7684\u521a\u6027\u53d8\u5316\uff08\u65cb\u8f6c\u6216\u5e73\u79fb\uff09\uff0c\u5750\u6807\u53d1\u751f\u53d8\u5316\u3002\u5176\u5b9e\u5bf9\u4e8e\u70b9\u4e91\u5206\u7c7bor\u5206\u5272\u4efb\u52a1\u6765\u8bf4\uff08\u5206\u5272\u53ef\u4ee5\u8ba4\u4e3a\u662f\u70b9\u7684\u5206\u7c7b\uff09\uff0c\u4f8b\u5982\uff0c\u6574\u4f53\u7684\u65cb\u8f6c\u548c\u5e73\u79fb\u4e0d\u5e94\u4fee\u6539\u5168\u5c40\u70b9\u4e91\u7c7b\u522b\u548c\u6bcf\u4e2a\u70b9\u7684\u7c7b\u522b<\/strong>\uff0c\u4e5f\u4e0d\u5e94\u4fee\u6539\u70b9\u7684\u5206\u5272\u56e0\u6b64\u9700\u8981\u4fdd\u8bc1\u4eff\u5c04\u53d8\u6362\u65e0\u5173\u6027<\/span>\uff08\u7b80\u5355\u6765\u8bf4\uff0c\u201c\u4eff\u5c04\u53d8\u6362\u201d\u5c31\u662f\uff1a\u201c\u7ebf\u6027\u53d8\u6362\u201d+\u201c\u5e73\u79fb\u201d\uff09\uff0c\u4f46\u662f\u5bf9\u4e8e\u4f4d\u7f6e \u654f\u611f\u7684 \u4efb\u52a1\uff1a\u70b9\u4e91\u914d\u51c6\u3001\u70b9\u4e91\u8865\u5168\u4efb\u52a1\uff0c\u5bf9\u4e8e\u4f4d\u7f6e\u654f\u611f\uff0c\u5c31\u4e0d\u9700\u8981\u4fdd\u8bc1 \u4eff\u5c04\u53d8\u6362\u7684\u65e0\u5173\u6027<\/strong><\/span> \u3002<\/p>\n\n\n\n

      \u6211\u4eec\u5e0c\u671b\u4e0d\u8bba\u70b9\u4e91\u5728\u600e\u6837\u7684\u5750\u6807\u7cfb\u4e0b\u5448\u73b0\uff0c\u7f51\u7edc\u90fd\u80fd\u6b63\u786e\u7684\u8bc6\u522b\u51fa\u3002\u8fd9\u4e2a\u95ee\u9898\u53ef\u4ee5\u901a\u8fc7STN\uff08spacial transform netw\uff09\u6765\u89e3\u51b3\u3002\u4e09\u7ef4\u4e0d\u592a\u4e00\u6837\u7684\u662f\u70b9\u4e91\u662f\u4e00\u4e2a\u4e0d\u89c4\u5219\u7684\u7ed3\u6784\uff08\u65e0\u5e8f\uff0c\u65e0\u7f51\u683c\uff09\uff0c\u4e0d\u9700\u8981\u91cd\u91c7\u6837\u7684\u8fc7\u7a0b\u3002pointnet\u901a\u8fc7\u5b66\u4e60\u4e00\u4e2a\u77e9\u9635\u6765\u8fbe\u5230\u5bf9\u76ee\u6807\u6700\u6709\u6548\u7684\u53d8\u6362\u3002<\/p>\n\n\n\n

      \u89e3\u51b3\u65b9\u6cd5 <\/p>\n\n\n\n

      1. \u7a7a\u95f4\u53d8\u6362\u7f51\u7edc\u89e3\u51b3\u65cb\u8f6c\u95ee\u9898\uff1a\u4e09\u7ef4\u7684STN\u53ef\u4ee5\u901a\u8fc7\u5b66\u4e60\u70b9\u4e91\u672c\u8eab\u7684\u4f4d\u59ff\u4fe1\u606f\u5b66\u4e60\u5230\u4e00\u4e2a\u6700\u6709\u5229\u4e8e\u7f51\u7edc\u8fdb\u884c\u5206\u7c7b\u6216\u5206\u5272\u7684DxD\u65cb\u8f6c\u77e9\u9635\uff08D\u4ee3\u8868\u7279\u5f81\u7ef4\u5ea6\uff0cpointnet\u4e2dD\u91c7\u75283\u548c64\uff09\u3002\u81f3\u4e8e\u5176\u4e2d\u7684\u539f\u7406\uff0c\u6211\u7684\u7406\u89e3\u662f\uff0c\u901a\u8fc7\u63a7\u5236\u6700\u540e\u7684loss\u6765\u5bf9\u53d8\u6362\u77e9\u9635\u8fdb\u884c\u8c03\u6574\uff0cpointnet\u5e76\u4e0d\u5173\u5fc3\u6700\u540e\u771f\u6b63\u505a\u4e86\u4ec0\u4e48\u53d8\u6362\uff0c\u53ea\u8981\u6709\u5229\u4e8e\u6700\u540e\u7684\u7ed3\u679c\u90fd\u53ef\u4ee5\u3002pointnet\u91c7\u7528\u4e86\u4e24\u6b21STN\uff0c\u7b2c\u4e00\u6b21input transform\u662f\u5bf9\u7a7a\u95f4\u4e2d\u70b9\u4e91\u8fdb\u884c\u8c03\u6574\uff0c\u76f4\u89c2\u4e0a\u7406\u89e3\u662f\u65cb\u8f6c\u51fa\u4e00\u4e2a\u66f4\u6709\u5229\u4e8e\u5206\u7c7b\u6216\u5206\u5272\u7684\u89d2\u5ea6\uff0c\u6bd4\u5982\u628a\u7269\u4f53\u8f6c\u5230\u6b63\u9762\uff1b\u7b2c\u4e8c\u6b21feature transform\u662f\u5bf9\u63d0\u53d6\u51fa\u768464\u7ef4\u7279\u5f81\u8fdb\u884c\u5bf9\u9f50\uff0c\u5373\u5728\u7279\u5f81\u5c42\u9762\u5bf9\u70b9\u4e91\u8fdb\u884c\u53d8\u6362\u3002<\/li>
      2. maxpooling\u89e3\u51b3\u65e0\u5e8f\u6027\u95ee\u9898\uff1a\u7f51\u7edc\u5bf9\u6bcf\u4e2a\u70b9\u8fdb\u884c\u4e86\u4e00\u5b9a\u7a0b\u5ea6\u7684\u7279\u5f81\u63d0\u53d6\u4e4b\u540e\uff0cmaxpooling\u53ef\u4ee5\u5bf9\u70b9\u4e91\u7684\u6574\u4f53\u63d0\u53d6\u51faglobal feature\u3002<\/li><\/ol>\n\n\n\n

        3.2 \u7f51\u7edc\u7ed3\u6784:<\/p>\n\n\n\n

        \"\"
        batchnormal\u5bf9\u4e8e\u4e0a\u91c7\u6837\u4efb\u52a1\u6765\u8bf4\u6548\u679c\u4e0d\u597d<\/figcaption><\/figure>\n\n\n\n

        \u7f51\u7edc\u5206\u6210\u4e86\u5206\u7c7b\u7f51\u7edc\u548c\u5206\u5272\u7f51\u7edc2\u4e2a\u90e8\u5206\uff0c\u5927\u4f53\u601d\u8def\u7c7b\u4f3c\uff0c\u90fd\u662f\u8bbe\u8ba1\u8868\u5f81\u7684\u8fc7\u7a0b\u5206\u7c7b\u7f51\u7edc\u8bbe\u8ba1global feature\uff0c\u5206\u5272\u7f51\u7edc\u8bbe\u8ba1point-wise feature\u3002\u4e24\u8005\u90fd\u662f\u4e3a\u4e86\u8ba9\u8868\u5f81\u5c3d\u53ef\u80fddiscriminative\uff0c\u4e5f\u5c31\u662f\u540c\u7c7b\u7684\u80fd\u5206\u5230\u4e00\u7c7b\uff0c\u4e0d\u540c\u7c7b\u7684\u8ddd\u79bb\u80fd\u62c9\u5f00\u3002<\/p>\n\n\n\n

        \u8f93\u5165 n*3 n\u662f\u70b9\u6570<\/p>\n\n\n\n

        inputtransform\uff1a\u653e\u5c04\u53d8\u6362\uff08\u4e3a\u4e86\u4fdd\u8bc1\u4eff\u5c04\u53d8\u6362\u7684\u4e0d\u53d8\u6027\uff09\uff1a\u76f4\u63a5\u9884\u6d4b\u4e00\u4e2a\u53d8\u6362\u77e9\u9635\uff083*3\uff09\u6765\u5904\u7406\u8f93\u5165\u70b9\u7684\u5750\u6807\uff08\u5bf9\u6240\u6709\u5750\u6807\u8fdb\u884c\u53d8\u6362\uff09\u3002\u56e0\u4e3a\u4f1a\u6709\u6570\u636e\u589e\u5f3a\u7684\u64cd\u4f5c\u5b58\u5728\uff0c\u8fd9\u6837\u505a\u53ef\u4ee5\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u4fdd\u8bc1\u7f51\u7edc\u53ef\u4ee5\u5b66\u4e60\u5230\u53d8\u6362\u65e0\u5173\u6027<\/strong>\u3002T-Net\u6a21\u578b\uff0c\u5b83\u7684\u4e3b\u8981\u4f5c\u7528\u662f\u5b66\u4e60\u51fa\u53d8\u5316\u77e9\u9635\u6765\u5bf9\u8f93\u5165\u7684\u70b9\u4e91\u6216\u7279\u5f81<\/strong>\u8fdb\u884c\u89c4\u8303\u5316\u5904\u7406\u3002<\/p>\n\n\n\n

        MLP\uff1a<\/p>\n\n\n\n

        \u6709\u4e24\u79cd\u5b9e\u73b0\u65b9\u6cd5\uff1a<\/p>\n\n\n\n

        1\u3001\u8f93\u5165 B,N,3 —- nn.liner\u5c42 — B,N,64<\/p>\n\n\n\n

        2\u3001\u8f93\u5165 B,3,N —- conv1d(1×1) — B,64,N <\/p>\n\n\n\n

        Pooling:<\/p>\n\n\n\n

        \u4e3a\u4e86\u89e3\u51b3\u65e0\u5e8f\u6027\uff08\u70b9\u4e91\u672c\u8d28\u4e0a\u662f\u4e00\u957f\u4e32\u70b9\uff08nx3\u77e9\u9635\uff0c\u5176\u4e2dn\u662f\u70b9\u6570\uff09\u3002\u5728\u51e0\u4f55\u4e0a\uff0c\u70b9\u7684\u987a\u5e8f\u4e0d\u5f71\u54cd\u5b83\u5728\u7a7a\u95f4\u4e2d\u5bf9\u6574\u4f53\u5f62\u72b6\u7684\u8868\u793a\uff0c\u4f8b\u5982\uff0c\u76f8\u540c\u7684\u70b9\u4e91\u53ef\u4ee5\u7531\u4e24\u4e2a\u5b8c\u5168\u4e0d\u540c\u7684\u77e9\u9635\u8868\u793a\u3002\uff09<\/em>\u4f7f\u7528 maxpooling\u6216sumpooling\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u6700\u540e\u7684D\u7ef4\u7279\u5f81\u5bf9\u6bcf\u4e00\u7ef4\u90fd\u9009\u53d6N\u4e2a\u70b9\u4e2d\u5bf9\u5e94\u7684\u6700\u5927\u7279\u5f81\u503c\u6216\u7279\u5f81\u503c\u603b\u548c\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u901a\u8fc7g\u6765\u89e3\u51b3\u65e0\u5e8f\u6027\u95ee\u9898\u3002<\/p>\n\n\n\n

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

        \u6700\u540e\u518d\u7ecf\u8fc7\u4e00\u4e2amlp\uff08\u4ee3\u7801\u4e2d\u8fd0\u7528\u5168\u8fde\u63a5\uff09\u5f97\u5230k\u4e2ascore\u3002\u5206\u7c7b\u7f51\u7edc\u6700\u540e\u63a5\u7684loss\u662fsoftmax\u3002<\/p>\n\n\n\n

        \u5206\u5272\u7f51\u7edc\uff1a<\/p>\n\n\n\n

        \u5c06\u6c60\u5316\u540e\u7684\u7279\u5f81\u548c\u524d\u4e00\u9636\u6bb5\u7279\u5f81\u62fc\u63a5\uff0c\u6c60\u5316\u540e\u7684\u7279\u5f81\u6709\u5168\u5c40\u4fe1\u606f\uff0c\u5728\u548c\u4e4b\u524d\u7684\u62fc\u63a5\uff0c\u4ee5\u6b64\u5f97\u5230\u540c\u65f6\u5bf9\u5c40\u90e8\u4fe1\u606f\u548c\u5168\u5c40\u4fe1\u606f\u611f\u77e5\u7684point-wise\u7279\u5f81\uff0c\u63d0\u5347\u8868\u5f81\u6548\u679c\u3002\u7136\u540e\u6700\u540e\u8f93\u51fan*m, m\u4e3a\u7c7b\u522b\u6570\u91cf\uff0c\u8868\u793a\u6bcf\u4e2a\u70b9\u7684\u7c7b\u522b\u4fe1\u606f\u3002<\/p>\n\n\n\n

        \u635f\u5931\u51fd\u6570\uff1a<\/strong><\/p>\n\n\n\n

        \u5206\u7c7b\u4e2d\u5e38\u7528\u7684\u4ea4\u53c9\u71b5+alignment network\u4e2d\u7528\u4e8e\u7ea6\u675f\u751f\u6210\u7684alignment matrix\u7684loss<\/p>\n\n\n\n

        dataset and experiments<\/strong><\/h3>\n\n\n\n

        evaluate metric<\/strong><\/p>\n\n\n\n

        \u5206\u7c7b\uff1a\u5206\u7c7b\u51c6\u786e\u7387acc
        \u5206\u5272\uff1amIoU<\/p>\n\n\n\n

        dataset<\/strong><\/p>\n\n\n\n

        \u5206\u7c7b\uff1aModelNet40
        \u5206\u5272\uff1aShapeNet Part dataset\u548cStanford 3D semantic parsing dataset<\/p>\n\n\n\n

        experiments<\/strong><\/p>\n\n\n\n

        1\u3001\u5206\u7c7b\uff1a<\/p>\n\n\n\n

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

        2\u3001\u5c40\u90e8\u5206\u5272\uff1a<\/p>\n\n\n\n

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

        code\uff1a<\/h3>\n\n\n\n

        1. \u5982\u4f55\u5bf9\u70b9\u4e91\u4f7f\u7528MLP\uff1f
        2. alignment network\u600e\u4e48\u505a\u7684\uff1f
        3. \u5bf9\u79f0\u51fd\u6570\u5982\u4f55\u5b9e\u73b0\u6765\u63d0\u53d6global feature\u7684\uff1f
        4. loss?<\/p>\n\n\n\n

        def get_model(point_cloud, is_training, bn_decay=None):\n    \"\"\" Classification PointNet, input is BxNx3, output Bx40 \"\"\"\n    batch_size = point_cloud.get_shape()[0].value\n    num_point = point_cloud.get_shape()[1].value\n    end_points = {}\n    with tf.variable_scope('transform_net1') as sc:\n        transform = input_transform_net(point_cloud, is_training, bn_decay, K=3)\n    point_cloud_transformed = tf.matmul(point_cloud, transform)\n    input_image = tf.expand_dims(point_cloud_transformed, -1)\n    net = tf_util.conv2d(input_image, 64, [1,3],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='conv1', bn_decay=bn_decay)\n    net = tf_util.conv2d(net, 64, [1,1],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='conv2', bn_decay=bn_decay)\n    with tf.variable_scope('transform_net2') as sc:\n        transform = feature_transform_net(net, is_training, bn_decay, K=64)\n    end_points['transform'] = transform\n    net_transformed = tf.matmul(tf.squeeze(net, axis=[2]), transform)\n    net_transformed = tf.expand_dims(net_transformed, [2])\n    net = tf_util.conv2d(net_transformed, 64, [1,1],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='conv3', bn_decay=bn_decay)\n    net = tf_util.conv2d(net, 128, [1,1],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='conv4', bn_decay=bn_decay)\n    net = tf_util.conv2d(net, 1024, [1,1],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='conv5', bn_decay=bn_decay)\n    # Symmetric function: max pooling<\/em>\n    net = tf_util.max_pool2d(net, [num_point,1],\n                             padding='VALID', scope='maxpool')\n    net = tf.reshape(net, [batch_size, -1])\n    net = tf_util.fully_connected(net, 512, bn=True, is_training=is_training,\n                                  scope='fc1', bn_decay=bn_decay)\n    net = tf_util.dropout(net, keep_prob=0.7, is_training=is_training,\n                          scope='dp1')\n    net = tf_util.fully_connected(net, 256, bn=True, is_training=is_training,\n                                  scope='fc2', bn_decay=bn_decay)\n    net = tf_util.dropout(net, keep_prob=0.7, is_training=is_training,\n                          scope='dp2')\n    net = tf_util.fully_connected(net, 40, activation_fn=None, scope='fc3')\n    return net, end_points<\/code><\/pre>\n\n\n\n
        MLP\u7684\u6838\u5fc3\u505a\u6cd5\uff1a<\/p>\n\n\n\n
        input_image = tf.expand_dims(point_cloud_transformed, -1)\nnet = tf_util.conv2d(input_image, 64, [1,3],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='conv1', bn_decay=bn_decay)\nnet = tf_util.conv2d(net, 64, [1,1],\n                     padding='VALID', stride=[1,1],\n                     bn=True, is_training=is_training,\n                     scope='conv2', bn_decay=bn_decay)\n<\/code><\/pre>\n\n\n\n
        \u8fd9\u91ccinput_image\u7ef4\u5ea6\u662f B\u00d7N\u00d73\u00d71 \uff0c\u56e0\u6b64\u5c06\u70b9\u4e91\u770b\u6210\u662fW\u548cH\u5206\u4e3aN\u548c3\u76842D\u56fe\u50cf\uff0c\u7ef4\u5ea6\u662f 1<\/p>\n\n\n\n
        \u7136\u540e\u76f4\u63a5\u57fa\u4e8e\u8fd9\u4e2a\u201c2D\u56fe\u50cf\u201d\u505a\u5377\u79ef\uff0c\u7b2c\u4e00\u4e2a\u5377\u79ef\u6838size\u662f [1,3] \uff0c\u6b63\u597d\u5bf9\u5e94\u7684\u5c31\u662f\u201c2D\u56fe\u50cf\u201d\u7684\u4e00\u884c\uff0c\u4e5f\u5c31\u662f\u4e00\u4e2a\u70b9\uff08\u4e09\u7ef4\u5750\u6807\uff09\uff0c\u8f93\u51fa\u901a\u9053\u6570\u662f64\uff0c\u56e0\u6b64\u8f93\u51fa\u5f20\u91cf\u7ef4\u5ea6\u5e94\u8be5\u662f B\u00d7N\u00d71\u00d764<\/p>\n\n\n\n
        \u7b2c\u4e8c\u4e2a\u5377\u79ef\u6838size\u662f [1,1] \uff0c 1\u22171 \u5377\u79ef\u53ea\u6539\u53d8\u901a\u9053\u6570\uff0c\u8f93\u51fa\u5f20\u91cf\u7ef4\u5ea6\u662f B\u00d7N\u00d71\u00d764<\/p>\n\n\n\n
        conv2d\u5c31\u662f\u5c06\u5377\u79ef\u5c01\u88c5\u4e86\u4e00\u4e0b\uff0c\u6838\u5fc3\u90e8\u5206\u4e5f\u5c31\u662f\u8c03\u7528tf.nn.conv2d\uff0c\u5b9e\u73b0\u5982\u4e0b\uff1a<\/p>\n\n\n\n
        def conv2d(inputs,\n           num_output_channels,\n           kernel_size,\n           scope,\n           stride=[1, 1],\n           padding='SAME',\n           use_xavier=True,\n           stddev=1e-3,\n           weight_decay=0.0,\n           activation_fn=tf.nn.relu,\n           bn=False,\n           bn_decay=None,\n           is_training=None):\n  \"\"\" 2D convolution with non-linear operation.\n  Args:\n    inputs: 4-D tensor variable BxHxWxC\n    num_output_channels: int\n    kernel_size: a list of 2 ints\n    scope: string\n    stride: a list of 2 ints\n    padding: 'SAME' or 'VALID'\n    use_xavier: bool, use xavier_initializer if true\n    stddev: float, stddev for truncated_normal init\n    weight_decay: float\n    activation_fn: function\n    bn: bool, whether to use batch norm\n    bn_decay: float or float tensor variable in [0,1]\n    is_training: bool Tensor variable\n  Returns:\n    Variable tensor\n  \"\"\"\n  with tf.variable_scope(scope) as sc:\n      kernel_h, kernel_w = kernel_size\n      num_in_channels = inputs.get_shape()[-1].value\n      kernel_shape = [kernel_h, kernel_w,\n                      num_in_channels, num_output_channels]\n      kernel = _variable_with_weight_decay('weights',\n                                           shape=kernel_shape,\n                                           use_xavier=use_xavier,\n                                           stddev=stddev,\n                                           wd=weight_decay)\n      stride_h, stride_w = stride\n      outputs = tf.nn.conv2d(inputs, kernel,\n                             [1, stride_h, stride_w, 1],\n                             padding=padding)\n      biases = _variable_on_cpu('biases', [num_output_channels],\n                                tf.constant_initializer(0.0))\n      outputs = tf.nn.bias_add(outputs, biases)\n      if bn:\n        outputs = batch_norm_for_conv2d(outputs, is_training,\n                                        bn_decay=bn_decay, scope='bn')\n      if activation_fn is not None:\n        outputs = activation_fn(outputs)\n      return outputs<\/code><\/pre>\n\n\n\n
         alignment network \uff1a<\/p>\n\n\n\n
        input_transform_net\u4e3a\u4f8b\uff1a<\/p>\n\n\n\n
        def input_transform_net(point_cloud, is_training, bn_decay=None, K=3):\n    \"\"\" Input (XYZ) Transform Net, input is BxNx3 gray image\n        Return:\n            Transformation matrix of size 3xK \"\"\"\n    batch_size = point_cloud.get_shape()[0].value\n    num_point = point_cloud.get_shape()[1].value\n    input_image = tf.expand_dims(point_cloud, -1)\n    net = tf_util.conv2d(input_image, 64, [1,3],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='tconv1', bn_decay=bn_decay)\n    net = tf_util.conv2d(net, 128, [1,1],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='tconv2', bn_decay=bn_decay)\n    net = tf_util.conv2d(net, 1024, [1,1],\n                         padding='VALID', stride=[1,1],\n                         bn=True, is_training=is_training,\n                         scope='tconv3', bn_decay=bn_decay)\n    net = tf_util.max_pool2d(net, [num_point,1],\n                             padding='VALID', scope='tmaxpool')\n    net = tf.reshape(net, [batch_size, -1])\n    net = tf_util.fully_connected(net, 512, bn=True, is_training=is_training,\n                                  scope='tfc1', bn_decay=bn_decay)\n    net = tf_util.fully_connected(net, 256, bn=True, is_training=is_training,\n                                  scope='tfc2', bn_decay=bn_decay)\n    with tf.variable_scope('transform_XYZ') as sc:\n        assert(K==3)\n        weights = tf.get_variable('weights', [256, 3*K],\n                                  initializer=tf.constant_initializer(0.0),\n                                  dtype=tf.float32)\n        biases = tf.get_variable('biases', [3*K],\n                                 initializer=tf.constant_initializer(0.0),\n                                 dtype=tf.float32)\n        biases += tf.constant([1,0,0,0,1,0,0,0,1], dtype=tf.float32)\n        transform = tf.matmul(net, weights)\n        transform = tf.nn.bias_add(transform, biases)\n    transform = tf.reshape(transform, [batch_size, 3, K])\n    return transform<\/code><\/pre>\n\n\n\n
        \u5b9e\u9645\u4e0a\uff0c\u524d\u534a\u90e8\u5206\u5c31\u662f\u901a\u8fc7\u5377\u79ef\u548cmax_pooling\u5bf9batch\u5185\u5404\u4e2a\u70b9\u4e91\u63d0\u53d6global feature\uff0c\u518d\u5c06global feature\u964d\u5230 3\u00d7K \u7ef4\u5ea6\uff0c\u5e76reshape\u6210 3\u00d73 \uff0c\u5f97\u5230transform matrix<\/p>\n\n\n\n
        \u901a\u8fc7\u6570\u636e\u589e\u5f3a\u4e30\u5bcc\u8bad\u7ec3\u6570\u636e\u96c6\uff0c\u7f51\u7edc\u786e\u5b9e\u5e94\u8be5\u5b66\u4e60\u5230\u6709\u6548\u7684transform matrix\uff0c\u7528\u6765\u5b9e\u73b0transformation invariance<\/p>\n\n\n\n
        loss<\/strong>\uff1a<\/p>\n\n\n\n
        \u76d1\u7763\u5206\u7c7b\u4efb\u52a1\u4e2d\u5e38\u7528\u7684\u4ea4\u53c9\u71b5loss + alignment network\u4e2d\u7684mat_diff_loss<\/p>\n\n\n\n
        \"\"<\/figure><\/div>\n\n\n\n
         \u5bf9\u4e8e\u7279\u5f81\u7a7a\u95f4\u7684alignment network\uff0c\u7531\u4e8e\u7279\u5f81\u7a7a\u95f4\u7ef4\u5ea6\u6bd4\u8f83\u9ad8\uff0c\u56e0\u6b64\u76f4\u63a5\u751f\u6210\u7684alignment matrix\u4f1a\u7ef4\u5ea6\u7279\u522b\u5927\uff0c\u4e0d\u597d\u4f18\u5316\uff0c\u56e0\u6b64\u8fd9\u91cc\u9700\u8981\u52a0\u4e2aloss\u7ea6\u675f\u4e00\u4e0b\u3002<\/p>\n\n\n\n
        \u603b\u7ed3\uff1a<\/strong><\/p>\n\n\n\n
              PointNet\u4e4b\u6240\u4ee5\u5f71\u54cd\u529b\u5de8\u5927\uff0c\u5e76\u4e0d\u4ec5\u4ec5\u662f\u56e0\u4e3a\u5b83\u662f\u7b2c\u4e00\u7bc7\uff0c\u66f4\u91cd\u8981\u7684\u662f\u5b83\u7684\u7f51\u7edc\u5f88\u7b80\u6d01\uff08\u7b80\u6d01\u4e2d\u8574\u542b\u4e86\u5927\u91cf\u7684\u5de5\u4f5c\u6765\u63a2\u5bfb\u51fa\u7b80\u6d01\u8fd9\u6761\u8def\uff09\u5374\u975e\u5e38\u7684work\uff0c\u8fd9\u4e5f\u5c31\u4f7f\u5f97\u5b83\u80fd\u591f\u6210\u4e3a\u4e00\u4e2a\u5de5\u5177\uff0c\u4e00\u4e2a\u4e3a\u70b9\u4e91\u8868\u5f81\u7684encoder\u5de5\u5177\uff0c\u5e94\u7528\u5230\u66f4\u5e7f\u9614\u7684\u70b9\u4e91\u5904\u7406\u4efb\u52a1\u4e2d\u3002<\/p>\n\n\n\n
             MLP+max pooling\u7adf\u7136\u5c31\u51fb\u8d25\u4e86\u4f17\u591aSOTA\uff0c\u4ee4\u4eba\u60ca\u8bb6\u3002\u53e6\u5916PointNet\u5728\u4f17\u591a\u7ec6\u8282\u8bbe\u8ba1\u4e5f\u90fd\u8fdb\u884c\u4e86\u7406\u8bba\u5206\u6790\u548c\u6d88\u878d\u5b9e\u9a8c\u9a8c\u8bc1\uff0c\u4fdd\u8bc1\u4e86\u4e25\u8c28\u6027\uff0c\u8fd9\u4e5f\u4e3aPointNet\u540e\u9762\u80fd\u591f\u5927\u89c4\u6a21\u88ab\u5e94\u7528\u63d0\u4f9b\u4e86\u652f\u6301\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
        \u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/1612.00593\uff08cvpr2017\uff09 code\uff1ahttp … \u7ee7\u7eed\u9605\u8bfbpointnet–\u57fa\u4e8e\u70b9\u4e91\u7684\u5206\u7c7b\u548c\u5206\u5272\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,31,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/9021"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=9021"}],"version-history":[{"count":99,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/9021\/revisions"}],"predecessor-version":[{"id":9345,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/9021\/revisions\/9345"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=9021"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=9021"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=9021"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}