{"id":437375,"date":"2021-04-18T15:57:30","date_gmt":"2021-04-18T07:57:30","guid":{"rendered":"http:\/\/4563.org\/?p=437375"},"modified":"2021-04-18T15:57:30","modified_gmt":"2021-04-18T07:57:30","slug":"%e7%9b%96%e6%a5%bc%e6%8a%bd%e5%a5%96%ef%bd%9c-5-%e5%88%86%e9%92%9f%e5%ad%a6%e4%b9%a0%e9%98%bf%e9%87%8c%e7%ae%97%e6%b3%95%e5%86%a0%e5%86%9b%e6%96%b9%e6%a1%88","status":"publish","type":"post","link":"http:\/\/4563.org\/?p=437375","title":{"rendered":"\u76d6\u697c\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848"},"content":{"rendered":"<div>\n<div>\n<div>\n<h1>                  \u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848               <\/h1>\n<p> <\/p>\n<div>\n<div> <span>\u8cc7\u6df1\u5927\u4f6c : VideoCloudTech <\/span>  <span><i><\/i> 0<\/span> <\/div>\n<div> <\/div>\n<\/p><\/div>\n<\/p><\/div>\n<\/p><\/div>\n<div isfirst=\"1\"> <\/p>\n<blockquote>\n<p>\u7ecf\u5178\u7b97\u6cd5\u6848\u4f8b\u5206\u4eab\uff0c\u6587\u672b\u66f4\u6709\u76d6\u62bd\u5956\u6d3b\u52a8\uff01\u901f\u901f\u52a0\u5165\u963f\u91cc\u7b97\u6cd5\u4ea4\u6d41\u7fa4\uff01<\/p>\n<\/blockquote>\n<p>\u89c6\u9891\u76ee\u6807\u5206\u5272\u662f\u76ee\u524d\u89c6\u9891\u7b97\u6cd5\u9886\u57df\u7684\u524d\u6cbf\u8bdd\u9898\uff0c\u8d8a\u6765\u8d8a\u591a\u7684\u5e94\u7528\u5728\u4e86\u5404\u4e2a\u884c\u4e1a\u573a\u666f\u91cc\u3002\u672c\u6b21\u7531\u82f1\u7279\u5c14\u4e0e\u963f\u91cc\u4e91\u8054\u5408\u4e3e\u529e\u3001\u4e0e\u4f18\u9177\u6218\u7565\u5408\u4f5c\u7684 \u201c\u65b0\u5185\u5bb9 \u65b0\u4ea4\u4e92 \u201c\u5168\u7403\u89c6\u9891\u4e91\u521b\u65b0\u6311\u6218\u8d5b\u7b97\u6cd5\u8d5b\u9053\uff0c\u4e5f\u5c06\u76ee\u5149\u805a\u7126\u4e8e\u8fd9\u4e00\u4e2a\u9886\u57df\u3002\u5927\u8d5b\u81ea\u5f00\u8d5b\u4ee5\u6765\uff0c\u5df2\u7ecf\u5438\u5f15\u4e86 2000 \u591a\u652f\u961f\u4f0d\u53c2\u8d5b\uff0c\u6c47\u805a\u4e86\u5168\u7403\u7b97\u6cd5\u7cbe\u82f1\u3002<\/p>\n<p>\u672c\u6587\u5c06\u4ee5 \u201cMedia AI \u963f\u91cc\u5df4\u5df4\u6587\u5a31\u7b97\u6cd5\u6311\u6218\u8d5b\u201d \u4e3a\u4f8b\uff0c\u7cbe\u9009\u51fa\u7531\u4f18\u9177\u4eba\u5de5\u667a\u80fd\u90e8\u7b97\u6cd5\u56e2\u961f\u63d0\u51fa\u7684\u51a0\u519b\u65b9\u6848\uff0c\u4e3a\u672c\u5c4a\u5927\u8d5b\u9009\u624b\u63d0\u4f9b\u6210\u529f\u7684\u53c2\u8d5b\u7ecf\u9a8c\u548c\u6280\u672f\u89e3\u8bfb\u3002<\/p>\n<p>\u9664\u4f20\u7edf\u5206\u5272\u7b97\u6cd5\u9700\u8981\u89e3\u51b3\u7684\u89c6\u89d2\u5149\u7167\u53d8\u5316\u3001\u76ee\u6807\u5c3a\u5ea6\u53d8\u5316\u3001\u76ee\u6807\u95f4\u906e\u6321\u7b49\u96be\u70b9\u4e4b\u5916\uff0c\u9762\u5411\u89c6\u9891\u667a\u80fd\u751f\u4ea7\u7684\u4eba\u7269\u5206\u5272\u7b97\u6cd5\u8fd8\u9700\u8981\u5173\u6ce8\u5982\u4e0b\u7279\u6b8a\u96be\u70b9\uff1a<\/p>\n<ul>\n<li>\u89c6\u9891\u573a\u666f\u5185\u5bb9\u4e30\u5bcc\u591a\u6837\uff1a\u8981\u6c42\u7b97\u6cd5\u5728\u590d\u6742\u80cc\u666f\u5e72\u6270\u4e0b\u6b63\u786e\u53d1\u6398\u573a\u666f\u663e\u8457\u4e3b\u89d2<\/li>\n<li>\u590d\u6742\u8863\u7740 \/ \u624b\u6301 \/ \u9644\u5c5e\u7269\uff1a\u8981\u6c42\u7b97\u6cd5\u5145\u5206\u63cf\u8ff0\u76ee\u6807\u4e30\u5bcc\u548c\u590d\u6742\u7684\u5916\u89c2\u8bed\u4e49<\/li>\n<li>\u76ee\u6807\u4eba\u7269\u5feb\u901f\u5267\u70c8\u52a8\u4f5c\uff1a\u8981\u6c42\u7b97\u6cd5\u89e3\u51b3\u8fd0\u52a8\u6a21\u7cca\u3001\u5267\u70c8\u5f62\u53d8\u5e26\u6765\u7684\u8bef\u5206\u5272\u3001\u6f0f\u5206\u5272<\/li>\n<\/ul>\n<p>Media AI \u5927\u8d5b\u7684\u6570\u636e\u96c6\u9762\u5411\u9ad8\u7cbe\u5ea6 &#8211; \u5b9e\u4f8b\u7ea7 &#8211; \u89c6\u9891\u4eba\u7269\u5206\u5272\u4efb\u52a1\uff0c\u63d0\u4f9b\u4e86 1700 \u6bb5\u9010\u5e27\u7cbe\u7ec6\u6807\u6ce8\u89c6\u9891\u6570\u636e\uff0c\u5176\u4e2d\u521d\u8d5b\u548c\u590d\u8d5b\u5404 800 \u6bb5\u8bad\u7ec3\u96c6\u548c 50 \u6bb5\u6d4b\u8bd5\u96c6\uff0c\u5bf9\u4e0a\u8ff0\u96be\u70b9\u90fd\u6709\u4f53\u73b0\u3002 \u76f8\u6bd4\u4e8e\u5b66\u672f \/ \u5de5\u4e1a\u754c\u6807\u51c6\u7684 DAVIS \u548c YouTube-VOS \u6570\u636e\u96c6\uff0c\u672c\u6570\u636e\u96c6\u542b\u6709\u4e1a\u754c\u6700\u591a\u4eba\u4f53\u76ee\u6807\u6807\u6ce8\u6807\u7b7e\uff08 18 \u4e07\u5e27\uff0c30 \u4e07\u6807\u6ce8\u4eba\u4f53\u76ee\u6807\uff09\uff0c\u4e14\u5728\u6807\u6ce8\u7cbe\u5ea6\u3001\u5185\u5bb9\u5e7f\u5ea6\u7b49\u65b9\u9762\u5747\u5904\u4e8e\u884c\u4e1a\u9886\u5148\u5730\u4f4d\u3002\u8fd9\u6b21\u6570\u636e\u5e93\u4f9d\u6258\u4f18\u9177\u7ad9\u5185\u6d77\u91cf\u8d44\u6e90\uff0c\u56ca\u62ec\u53e4\u88c5\u5267\u96c6\u3001\u73b0\u4ee3\u5267\u96c6\u3001\u8857\u62cd\u3001\u8857\u821e\u3001\u8fd0\u52a8\uff08\u8dd1\u9177\u3001\u7403\u7c7b\u3001\u5065\u8eab\uff09\u7b49\u4e30\u5bcc\u5185\u5bb9\uff0c\u66f4\u52a0\u7b26\u5408\u667a\u80fd\u5316\u89c6\u9891\u5185\u5bb9\u751f\u4ea7\u7684\u73b0\u5b9e\u9700\u6c42\u3002\u7ecf\u8fc7\u7cbe\u7ec6\u5316\u4eba\u5de5\u6807\u6ce8\uff0c\u6570\u636e\u96c6\u771f\u503c\u63a9\u7801\u56fe\u7cbe\u51c6\u52fe\u52d2\u89c6\u9891\u4eba\u7269\u7684\u8fb9\u7f18\u7ec6\u8282\uff0c\u4e3a\u8bad\u7ec3\u4ee5\u53ca\u6d4b\u8bd5\u5206\u5272\u7b97\u6cd5\u7684\u51c6\u786e\u6027\u548c\u7cbe\u786e\u6027\u63d0\u4f9b\u4e86\u4f9d\u636e\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u8be5\u6570\u636e\u96c6\u8fd8\u9488\u5bf9\u4eba\u7269\u624b\u6301\u7269 \/ \u9644\u5c5e\u7269\u8fdb\u884c\u4e86\u6807\u6ce8\uff0c\u6709\u52a9\u4e8e\u7b97\u6cd5\u5bf9\u4eba\u7269\u4e0e\u7269\u54c1\u4ece\u5c5e\u5173\u7cfb\u8fdb\u884c\u5b66\u4e60\u3002<\/p>\n<h2>\u51a0\u519b\u65b9\u6848\u7b97\u6cd5\u8be6\u89e3<\/h2>\n<p>\u5728\u521d\u8d5b\u9636\u6bb5\uff0c\u4f18\u9177\u4eba\u5de5\u667a\u80fd\u90e8\u7b97\u6cd5\u56e2\u961f\u4ee5 STM \uff08 Video Object Segmentation using Space-Time Memory Networks \uff09\u4e3a\u57fa\u7840\uff0c\u8fdb\u884c\u4e86\u5f7b\u5e95\u7684\u6a21\u578b\u590d\u73b0\u548c\u4ee5\u53ca\u7b97\u6cd5\u6539\u8fdb\u3002\u5728\u590d\u8d5b\u9636\u6bb5\uff0c\u4ee5\u521d\u8d5b\u534a\u76d1\u7763\u6a21\u578b\u4e3a\u9aa8\u5e72\uff0c\u914d\u5408\u4ee5\u76ee\u6807\u68c0\u6d4b\u3001\u663e\u8457\u6027\u5224\u522b\u3001\u5173\u952e\u5e27\u9009\u62e9\u7b49\u6a21\u5757\uff0c\u5b9e\u73b0\u9ad8\u7cbe\u5ea6\u65e0\u76d1\u7763\u89c6\u9891\u5206\u5272\u94fe\u8def\u3002<\/p>\n<h3>1. \u76d1\u7763\u89c6\u9891\u4eba\u7269\u5206\u5272<\/h3>\n<p>\u534a\u76d1\u7763 VOS \u7684\u4efb\u52a1\u76ee\u6807\u662f\u5728\u7ed9\u5b9a\u7b2c\u4e00\u5e27\u7269\u4f53\u63a9\u7801\u7684\u524d\u63d0\u4e0b\uff0c\u5c06\u8be5\u7269\u4f53\u5728\u540e\u7eed\u5e27\u4e2d\u8fde\u7eed\u5206\u5272\u51fa\u6765\u3002<\/p>\n<h4>1. 1 \u57fa\u672c\u6846\u67b6<\/h4>\n<ul>\n<li>\u63d0\u51fa Spatial Constrained Memory Reader \u4ee5\u89e3\u51b3 STM \u7a7a\u95f4\u8fde\u7eed\u6027\u4e0d\u8db3\u95ee\u9898<\/li>\n<\/ul>\n<p>\u9996\u5148 STM \u5728\u50cf\u7d20\u5339\u914d\u65f6\u662f\u57fa\u4e8e\u5916\u89c2\u7684\u5339\u914d\uff0c\u6ca1\u6709\u8003\u8651\u7269\u4f53\u5728\u76f8\u90bb\u4e24\u5e27\u4e4b\u95f4\u7a7a\u95f4\u4e0a\u7684\u8fde\u7eed\u6027\u3002\u6362\u8a00\u4e4b\uff0cSTM \u4f1a\u5bfb\u627e\u4e0e\u524d\u9762\u5e27\u4e2d\u5916\u89c2\u76f8\u4f3c\u7684\u7269\u4f53\uff0c\u4f46\u5bf9\u8be5\u7269\u4f53\u51fa\u73b0\u5728\u4f55\u4f4d\u7f6e\u4e0d\u654f\u611f\u3002\u56e0\u6b64\uff0c\u5f53\u4e00\u5e27\u4e2d\u51fa\u73b0\u591a\u4e2a\u5916\u89c2\u76f8\u4f3c\u7269\u4f53\u65f6\uff0cSTM \u7684\u5206\u5272\u7ed3\u679c\u5c31\u6709\u53ef\u80fd\u4ea7\u751f\u9519\u8bef\u3002<\/p>\n<p>\u9488\u5bf9\u8fd9\u4e2a\u95ee\u9898\uff0cDAVIS2020 \u534a\u76d1\u7763\u7b2c\u4e00\u540d\u65b9\u6848\u7684\u89e3\u51b3\u65b9\u6cd5\u662f\u5c06\u524d\u4e00\u5e27\u7684\u7269\u4f53 mask \u7ed3\u5408\u5230 encode \u4e4b\u540e\u7684 feature \u4e2d\uff0c\u964d\u4f4e\u79bb\u524d\u4e00\u5e27\u7269\u4f53\u4f4d\u7f6e\u8f83\u8fdc\u50cf\u7d20\u7684\u6743\u91cd\uff08\u5982\u56fe 1 \u6240\u793a\uff09\u3002\u5b9e\u9645\u5c1d\u8bd5\u540e\u53d1\u73b0\u589e\u76ca\u4e0d\u5927\u3002\u6211\u4eec\u8ba4\u4e3a\u539f\u56e0 \u5728\u4e8e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7ed9\u4e0e\u6a21\u578b\u8fc7\u5f3a\u7684\u4f4d\u7f6e\u5148\u9a8c\uff0c\u5bfc\u81f4\u6a21\u578b\u5206\u5272\u8fc7\u5206\u4f9d\u8d56\u4e8e\u524d\u4e00\u5e27\u7684\u7269\u4f53\u4f4d\u7f6e\u4fe1\u606f\uff0c\u7ea6\u675f\u4e86 non-local \u7684\u957f\u8ddd\u79bb\u5339\u914d\u80fd\u529b\u3002\u4e00\u65e6\u51fa\u73b0\u524d\u4e00\u5e27\u7269\u4f53\u88ab\u906e\u6321\uff0c\u6216\u8005\u524d\u4e00\u5e27\u7269\u4f53\u5206\u5272\u9519\u8bef\u7684\u60c5\u51b5\uff0c\u6574\u6bb5\u89c6\u9891\u7684\u5206\u5272\u7ed3\u679c\u5c06\u51fa\u73b0\u4e0d\u7a33\u5b9a\u6027\u504f\u79fb\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"http:\/\/4563.org\/wp-content\/uploads\/2021\/04\/20210419_607d3648e310a.png\" alt=\"\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848\" \/><\/p>\n<p>\u56fe 1. Spatial-contrained STM<\/p>\n<p>DAVIS2020 \u534a\u76d1\u7763\u7b2c\u4e09\u540d\u65b9\u6848\u5bf9\u6b64\u7684\u89e3\u51b3\u65b9\u6848\u662f kernelized memory reader \uff08\u5982\u56fe 2 \u6240\u793a\uff09\uff0c\u8fd9\u79cd\u65b9\u6cd5\u80fd\u4fdd\u8bc1 memory \u4e2d\u7684\u70b9\u4f1a\u5339\u914d\u5230 query \u4e2d\u6700\u76f8\u4f3c\u7684\u4e00\u4e2a\u533a\u57df\uff0c\u53ef\u4ee5\u907f\u514d\u51fa\u73b0\u4e00\u5bf9\u591a\u5339\u914d\u7684\u95ee\u9898\u3002\u4f46\u662f\u4e0d\u80fd\u4fdd\u8bc1\u7a7a\u95f4\u4e0a\u7684\u8fde\u7eed\uff0c\u5bb9\u6613\u51fa\u73b0\u4e0d\u53ef\u9006\u7684\u8bef\u5dee\u7d2f\u79ef\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"http:\/\/4563.org\/wp-content\/uploads\/2021\/04\/20210419_607d364f00697.png\" alt=\"\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848\" \/><\/p>\n<p>\u56fe 2. Kernelized-memory STM<\/p>\n<p>\u7efc\u5408\u8003\u8651\u4e0a\u8ff0\u4e24\u79cd\u65b9\u6cd5\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65e2\u80fd\u8003\u8651\u524d\u4e00\u5e27\u7269\u4f53\u4f4d\u7f6e\u4fe1\u606f\uff0c\u53c8\u4e0d\u5f71\u54cd\u539f\u59cb\u5339\u914d\u8bad\u7ec3\u8fc7\u7a0b\u7684\u65b9\u6cd5\u3002\u5982\u56fe 3 \u6240\u793a\uff0c\u5229\u7528\u524d\u4e00\u5e27\u7684\u7269\u4f53 mask \u751f\u6210\u4f4d\u7f6e\u5927\u5c0f\u76f8\u5173\u7684\u9ad8\u65af\u6838\uff0c\u901a\u8fc7\u8fd9\u4e2a\u9ad8\u65af\u6838\u6765\u4fee\u6b63 memory \u4e2d\u50cf\u7d20\u7684\u6700\u4f18\u5339\u914d\u4f4d\u7f6e\u3002\u4e4b\u540e\u6d41\u7a0b\u548c\u56fe 2 \u76f8\u4f3c\uff0c\u5229\u7528\u6bcf\u4e2a\u50cf\u7d20\u7684\u6700\u4f18\u5339\u914d\u4f4d\u7f6e\u5bf9\u539f\u59cb\u5339\u914d\u8fdb\u884c\u4fee\u6b63\u3002\u5982\u6b64\u4e00\u6765\uff0c\u65e2\u6ca1\u6709\u5f71\u54cd\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u5bfc\u81f4 non-local \u90e8\u5206\u5339\u914d\u80fd\u529b\u964d\u4f4e\uff0c\u53c8\u5f15\u5165\u4e86 spatial prior\uff0c\u4fdd\u8bc1\u4e86\u7269\u4f53 mask \u7684\u7a7a\u95f4\u8fde\u7eed\u6027\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"http:\/\/4563.org\/wp-content\/uploads\/2021\/04\/20210419_607d368e65296.png\" alt=\"\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848\" \/><\/p>\n<p>\u56fe 3.Spatial-contrained Mmeory Reader \u793a\u610f\u56fe<\/p>\n<ul>\n<li>\u589e\u52a0 ASPP &amp; HRNet post-refinement \u4ee5\u89e3\u51b3 STM \u89e3\u7801\u5668\u5bf9\u591a\u5c3a\u5ea6\u76ee\u6807\u5206\u5272\u7cbe\u7ec6\u5ea6\u8f83\u5dee\u95ee\u9898 \u901a\u8fc7 ASPP \u589e\u52a0\u591a\u5c3a\u5ea6\u4fe1\u606f\u7684\u6355\u83b7\u80fd\u529b\uff0c\u5229\u7528 HRnet \u5bf9 STM \u7684\u521d\u59cb\u5206\u5272\u7ed3\u679c\u8fdb\u884c refine\uff0c\u4f18\u5316\u7269\u4f53\u7ec6\u8282\u7684\u5206\u5272\u6548\u679c\u3002<\/li>\n<\/ul>\n<h4>1. 2 \u8bad\u7ec3\u7b56\u7565<\/h4>\n<p>\u5728\u6bd4\u8d5b\u4e2d\u91c7\u7528\u4e86\u4e24\u9636\u6bb5\u8bad\u7ec3\u7684\u65b9\u6cd5\u3002\u7b2c\u4e00\u4e2a\u9636\u6bb5\uff0c\u91c7\u7528 MS-COCO \u9759\u6001\u56fe\u50cf\u5e93\u6210\u89c6\u9891\u5e8f\u5217\u8fdb\u884c\u9884\u8bad\u7ec3\u3002\u7b2c\u4e8c\u4e2a\u9636\u6bb5\uff0c\u5c06\u516c\u5f00\u6570\u636e\u5e93\uff08 DAVIS,Youtube-vos \uff09\u548c\u6bd4\u8d5b\u8bad\u7ec3\u96c6\u8fdb\u884c\u5408\u5e76\u8bad\u7ec3\uff0c\u6765\u4fdd\u8bc1\u6709\u8db3\u591f\u7684\u6570\u636e\u91cf\u3002\u5177\u4f53\u8bad\u7ec3\u7ec6\u8282\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>Crop \u76f8\u90bb 3 \u5e27\u56fe\u50cf patch \u8fdb\u884c\u8bad\u7ec3\uff0c\u5c3d\u53ef\u80fd\u589e\u52a0 augmentation \u3002crop \u65f6\u9700\u8981\u6ce8\u610f\u4e00\u5b9a\u8981\u4fdd\u8bc1\u5728\u7b2c 2 \u548c\u7b2c 3 \u5e27\u51fa\u73b0\u7684\u7269\u4f53\u90fd\u5728\u7b2c\u4e00\u5e27\u51fa\u73b0\u4e86\uff0c\u5426\u5219\u5e94\u8be5\u8fc7\u6ee4\uff1b<\/li>\n<li>\u5c06 DAVIS,Youtube-vos \u548c\u6bd4\u8d5b\u8bad\u7ec3\u96c6\u4ee5\u4e00\u5b9a\u6bd4\u4f8b\u878d\u5408\u6548\u679c\u6700\u597d\uff1b<\/li>\n<li>\u8bad\u7ec3\u8fc7\u7a0b\u6307\u6807\u6ce2\u52a8\u8f83\u5927\uff0c\u91c7\u7528 poly \u5b66\u4e60\u7387\u7b56\u7565\u53ef\u7f13\u89e3\uff1b<\/li>\n<li>\u8bad\u7ec3\u6bd4\u8f83\u5403\u663e\u5b58\uff0cbatch size \u6bd4\u8f83\u5c0f\u7684\u8bdd\u8981 fix \u6240\u6709\u7684 bn \u5c42\u3002<\/li>\n<\/ul>\n<h4>1. 3 \u5176\u4ed6<\/h4>\n<p>\u2022 Backbone\uff1a\u66f4\u6362 resnest101 \u2022 \u6d4b\u8bd5\u7b56\u7565\uff1a\u4f7f\u7528 Multi-scale\/flip inference<\/p>\n<h4>1. 4 \u7ed3\u679c<\/h4>\n<p>\u4f18\u9177\u7b97\u6cd5\u56e2\u961f\u7684\u6a21\u578b\uff0c\u5728\u6d4b\u8bd5\u96c6\u4e0a\u53d6\u5f97\u4e86 95.5 \u7684\u6210\u7ee9\uff0c\u76f8\u6bd4\u539f\u59cb STM \u63d0\u9ad8\u5c06\u8fd1 5 \u4e2a\u70b9\u3002<\/p>\n<h3>2. \u65e0\u76d1\u7763\u89c6\u9891\u4eba\u7269\u5206\u5272<\/h3>\n<p>\u65e0\u76d1\u7763 VOS \u7684\u4efb\u52a1\u76ee\u6807\u662f\u5728\u4e0d\u7ed9\u5b9a\u4efb\u4f55\u6807\u6ce8\u4fe1\u606f\u7684\u524d\u63d0\u4e0b\uff0c\u81ea\u4e3b\u53d1\u6398\u524d\u666f\u76ee\u6807\u5e76\u8fdb\u884c\u8fde\u7eed\u7684\u5206\u5272\u3002\u65e0\u76d1\u7763 VOS \u65b9\u6cd5\u94fe\u8def\u8f83\u4e3a\u590d\u6742\uff0c\u901a\u5e38\u4e0d\u662f\u7531\u5355\u4e00\u6a21\u578b\u89e3\u51b3\uff0c\u5176\u4e2d\u6d89\u53ca\u5230\u76ee\u6807\u68c0\u6d4b\u3001\u6570\u636e\u5173\u8054\u3001\u8bed\u4e49\u5206\u5272\u3001\u5b9e\u4f8b\u5206\u5272\u7b49\u6a21\u5757\u3002<\/p>\n<h4>2.1 \u7b97\u6cd5\u6846\u67b6<\/h4>\n<p>\u6211\u4eec\u590d\u8d5b\u6240\u91c7\u7528\u7684\u7b97\u6cd5\u6d41\u7a0b\u5177\u4f53\u5206\u4e3a\u5982\u4e0b\u56db\u6b65\uff1a<\/p>\n<h4>a. \u9010\u5e27\u505a\u5b9e\u4f8b\u5206\u5272<\/h4>\n<p>\u91c7\u7528 DetectoRS \u4f5c\u4e3a\u68c0\u6d4b\u5668\uff0c\u4e3a\u4fdd\u8bc1\u6cdb\u5316\u80fd\u529b\uff0c\u6ca1\u6709\u5728\u6bd4\u8d5b\u8bad\u7ec3\u96c6\u4e0a finetune \u6a21\u578b\uff0c\u800c\u662f\u76f4\u63a5\u4f7f\u7528\u5728 MS COCO \u6570\u636e\u96c6\u8fdb\u884c\u8bad\u7ec3\u3002\u6b64\u9636\u6bb5\u53ea\u4fdd\u7559 person \u7c7b\u522b\u3002\u9608\u503c\u8bbe\u4e3a 0.1\uff0c\u76ee\u7684\u662f\u5c3d\u53ef\u80fd\u591a\u5730\u4fdd\u7559 proposal \u3002<\/p>\n<h4>b. \u5bf9\u5b9e\u4f8b\u5206\u5272\u7684 mask \u8fdb\u884c\u540e\u5904\u7406<\/h4>\n<p>\u5982\u4e0b\u5de6\u56fe\u6240\u793a\uff0c\u73b0\u6709 instance segmentation \u7684\u65b9\u6cd5\u4ea7\u751f\u7684 mask \u5206\u8fa8\u7387\u4f4e\uff0c\u8fb9\u7f18\u7c97\u7cd9\u3002\u6211\u4eec\u91c7\u7528\u8bed\u4e49\u5206\u5272\u6a21\u578b\u5bf9 DetectoRS \u4ea7\u751f\u7684\u7ed3\u679c\u8fdb\u884c refine \uff08 image+mask -&gt;HRnet -&gt; refined mask \uff09, \u7ed3\u679c\u5982\u4e0b\u56fe\u3002\u53ef\u4ee5\u770b\u51fa\u63a9\u7801\u56fe\u4e2d\u7684\u7269\u4f53\u8fb9\u7f18\u4ee5\u53ca\u7ec6\u8282\u90fd\u6709\u4e86\u660e\u663e\u7684\u6539\u5584\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"http:\/\/4563.org\/wp-content\/uploads\/2021\/04\/20210419_607d369486180.png\" alt=\"\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848\" \/><\/p>\n<p>\u56fe 4. \u68c0\u6d4b\u5668\uff08 DetectoRS \uff09\u8f93\u51fa\u63a9\u7801\u56fe\uff08\u4e0a\uff09\u4e0e refine \u540e\u63a9\u7801\u56fe\uff08\u4e0b\uff09<\/p>\n<h4>c. \u5e27\u95f4\u8fdb\u884c\u6570\u636e\u5173\u8054\uff0c\u5f97\u5230\u521d\u6b65\u7ed3\u679c<\/h4>\n<p>\u5229\u7528 STM \u5c06 t-1 \u5e27\u7684 mask warp \u5230 t \u5e27\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u5229\u7528 warp \u540e\u7684 mask \u548c\u7b2c t \u5e27\u7684\u5206\u5272\u7ed3\u679c\u8fdb\u884c\u5339\u914d\u3002\u901a\u8fc7\u8fd9\u4e2a\u8fc7\u7a0b\uff0c\u8865\u507f\u4e86\u8fd0\u52a8\u7b49\u95ee\u9898\u4ea7\u751f\u7684\u5f71\u54cd\uff0c\u7a33\u5b9a\u6027\u66f4\u9ad8\u3002\u5177\u4f53\u7684\uff0c\u5bf9\u4e8e\u9996\u5e27\u7269\u4f53\uff0c\u6211\u4eec\u4fdd\u7559\u7f6e\u4fe1\u5ea6\u5927\u4e8e 0.8 \u7684 proposal \u3002\u5bf9\u7b2c t-1 \u5e27\u548c\u7b2c t \u5e27\u505a\u6570\u636e\u5173\u8054\u65f6\uff0c\u9996\u5148\u5229\u7528 STM \u5c06\u7b2c t-1 \u5e27\u7684\u7ed3\u679c warp \u5230\u7b2c t \u5e27\u3002\u7136\u540e\u7528\u5308\u7259\u5229\u7b97\u6cd5\u5bf9 warp \u540e\u7684 mask \u548c\u7b2c t \u5e27\u7531 DetectRS \u751f\u6210\u7684 proposal \u8fdb\u884c\u4e8c\u90e8\u56fe\u5339\u914d\u3002<\/p>\n<h4>d. \u7b5b\u9009\u5206\u5272\u7ed3\u679c\u8f83\u597d\u7684\u5e27\u4f5c\u4e3a key frames \u8fdb\u884c\u8fed\u4ee3\u4f18\u5316<\/h4>\n<p>\u7ecf\u4e0a\u8ff0\u6570\u636e\u5173\u8054\u4ee5\u540e\uff0c\u6211\u4eec\u5df2\u7ecf\u5f97\u5230\u4e86\u521d\u6b65\u7684\u65e0\u76d1\u7763 VOS \u7ed3\u679c\uff0c\u5176\u4e2d\u6bcf\u5e27\u7684 mask \u662f\u7531 DetectRS \u751f\u6210\uff0cid \u662f\u7531\u6570\u636e\u5173\u8054\u51b3\u5b9a\u3002\u4f46\u662f\u8fd9\u4e2a\u7ed3\u679c\u5b58\u5728\u5f88\u591a\u95ee\u9898\uff0c\u8fd8\u53ef\u4ee5\u8fdb\u4e00\u6b65\u4f18\u5316\u3002\u6bd4\u5982\u8bf4\u89c6\u9891\u5f00\u59cb\u5904\u51fa\u73b0\u7684\u6f0f\u68c0\u65e0\u6cd5\u88ab\u8865\u4e0a\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5de6\u4fa7\u7684\u4eba\u5728\u89c6\u9891\u5f00\u59cb\u5904\u4e0d\u6613\u88ab\u68c0\u6d4b\uff0c\u76f4\u5230\u7b2c 10 \u5e27\u624d\u88ab\u68c0\u6d4b\u51fa\u6765\u3002\u53e6\u5916\uff0c\u89c6\u9891\u4e2d\u4eba\u4f53\u4ea4\u53e0\u4e25\u91cd\u5904\u5206\u5272\u8d28\u91cf\u8981\u8fdc\u4f4e\u4e8e\u4eba\u4f53\u8ddd\u79bb\u8f83\u5927\u5904\u3002 <img decoding=\"async\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"http:\/\/4563.org\/wp-content\/uploads\/2021\/04\/20210419_607d369a1c920.png\" alt=\"\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848\" \/><\/p>\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u53ef\u4ee5\u6839\u636e\u7269\u4f53\u6570\u91cf\uff0cbbox \u7684\u4ea4\u53e0\u7a0b\u5ea6\u7b49\u4fe1\u606f\u7b5b\u9009\u51fa\u4e00\u4e9b\u53ef\u80fd\u5206\u5272\u8f83\u597d\u7684\u5e27\u4f5c\u4e3a\u4e0b\u4e00\u8f6e\u4f18\u5316\u7684 reference \u3002\u5177\u4f53\u7684\uff0c\u6211\u4eec\u53ef\u4ee5\u5229\u7528\u7b5b\u9009\u51fa\u6765\u7684 key frames \u4f5c\u4e3a\u521d\u59cb memory\uff0c\u7528 STM \u8fdb\u884c\u53cc\u5411\u9884\u6d4b\u3002\u9996\u5148\u53cc\u5411\u9884\u6d4b\u53ef\u4ee5\u89e3\u51b3\u89c6\u9891\u5f00\u59cb\u5904\u7684\u6f0f\u68c0\uff0c\u5176\u6b21 STM \u5bf9\u4e8e\u906e\u6321\u7b49\u95ee\u9898\u7684\u5904\u7406\u4e5f\u8981\u597d\u8fc7\u5355\u5e27\u7684\u5b9e\u4f8b\u5206\u5272\u3002\u7ecf\u5b9e\u9a8c\u9a8c\u8bc1\uff0c\u6bcf\u8fed\u4ee3\u4e00\u6b21 STM \u53cc\u5411\u9884\u6d4b\uff0c\u6307\u6807\u90fd\u6709\u5c0f\u5e45\u5ea6\u63d0\u5347\u3002 <img decoding=\"async\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"http:\/\/4563.org\/wp-content\/uploads\/2021\/04\/20210419_607d36a06ea5e.png\" alt=\"\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848\" \/><\/p>\n<p>\u89c6\u9891\u76ee\u6807\uff08\u4eba\u7269\uff09\u5206\u5272\uff08 Video Object Segmentation\uff0c\u7b80\u79f0\u4e3a VOS \uff09\u7b97\u6cd5\u662f\u4e1a\u754c\u516c\u8ba4\u7684\u6280\u672f\u91cd\u70b9\u96be\u70b9\uff0c\u540c\u65f6\u53c8\u6709\u7740\u6700\u4e3a\u5e7f\u6cdb\u7684\u843d\u5730\u573a\u666f\u548c\u5e94\u7528\u9700\u6c42\u3002\u76f8\u4fe1\u53c2\u4e0e\u672c\u5c4a \u201c\u65b0\u5185\u5bb9 \u65b0\u4ea4\u4e92 \u201c\u5168\u7403\u89c6\u9891\u4e91\u521b\u65b0\u6311\u6218\u8d5b\u7b97\u6cd5\u8d5b\u9053\u7684\u9009\u624b\uff0c\u5c06\u4ee5\u89c6\u9891\u76ee\u6807\u5206\u5272\u4e3a\u8d77\u70b9\uff0c\u5229\u7528\u8ba1\u7b97\u673a\u89c6\u89c9\u7b97\u6cd5\u9886\u57df\u7684\u8bf8\u591a\u6280\u672f\uff0c\u4e3a\u884c\u4e1a\u548c\u5927\u4f17\u6253\u9020\u66f4\u52a0\u667a\u80fd\u5316\u3001\u4fbf\u6377\u5316\u3001\u8da3\u5473\u5316\u7684\u89c6\u9891\u670d\u52a1\u3002<\/p>\n<p><strong>\u5927\u5bb6\u597d\uff0c\u6211\u4eec\u662f\u963f\u91cc\u4e91\u89c6\u9891\u4e91\u56e2\u961f\u3002\u4e0a\u6b21\u6d3b\u52a8\u5df2\u7ecf\u63d0\u5230\u6700\u8fd1\u6211\u4eec\u5728\u4e3e\u529e\u5168\u7403\u89c6\u9891\u4e91\u521b\u65b0\u6311\u6218\u8d5b\uff0c\u4e0e\u6b64\u540c\u65f6\u6211\u4eec\u8fd8\u60f3\u628a\u4f18\u79c0\u7684\u7b97\u6cd5\u5206\u4eab\u7ed9\u5e7f\u5927\u5f00\u53d1\u8005\uff0c\u8fd9\u4e0d\uff0c\u6211\u4eec\u7acb\u9a6c\u7ec4\u7ec7\u4e86\u5305\u62ec\u963f\u91cc\u96c6\u56e2\u7cbe\u82f1\u7b97\u6cd5\u540c\u5b66\u5728\u5185\u7684\u6280\u672f\u4ea4\u6d41\u7fa4\u3002\u52a0\u5165\u6b64\u7fa4\uff0c\u4f60\u5c06\u83b7\u5f97\uff1a<\/strong><\/p>\n<ul>\n<li>\u963f\u91cc\u7b97\u6cd5\u5b66\u4e60\u8d44\u6599<\/li>\n<li>\u963f\u91cc\u51fa\u54c1\u514d\u8d39\u7b97\u6cd5\u8bfe\u7a0b<\/li>\n<li>\u7b97\u6cd5\u5927\u4f6c\u4e00\u5bf9\u4e00 battle<\/li>\n<li>\u6d3b\u8dc3\u7684\u6280\u672f\u8ba8\u8bba<\/li>\n<\/ul>\n<p><strong>\u6211\u4eec\u81f4\u529b\u4e8e\u6253\u9020\u6700\u9760\u8c31\u7b97\u6cd5\u6280\u672f\u4ea4\u6d41\u7fa4\uff0c\u9884\u544a\u4e00\u4e0b\uff0c\u6700\u8fd1\u6211\u4eec\u5c06\u63a8\u51fa\u7b97\u6cd5\u516c\u5f00\u8bfe\uff0c\u8bb2\u5e08\u662f\u6765\u81ea\u963f\u91cc\u4e91\u89c6\u9891\u4e91\u56e2\u961f\u7684\u5fb7\u56fd\u9ad8\u5bcc\u5e05\u535a\u58eb\uff0c\u5e26\u6765\u5174\u8da3\u533a\u57df\uff08 ROI \uff09\u667a\u80fd\u89c6\u9891\u7f16\u7801\u7684\u7b97\u6cd5\u5206\u4eab\u3002<\/strong><\/p>\n<p>\u5f53\u7136\uff0c\u73b0\u5728\u8fdb\u7fa4\u8fd8\u6709\u62bd\u5956\u6d3b\u52a8\uff01<\/p>\n<h2>\u5165\u7fa4\u62bd\u5956<\/h2>\n<ul>\n<li>\u5404\u4f4d\u5c0f\u4f19\u4f34\u626b\u7801\u5165\u7fa4<\/li>\n<li>\u9996\u6279\u5165\u7fa4\u5c0f\u4f19\u4f34\uff0c\u7ba1\u7406\u5458\u5c06\u62bd\u5956\u5f62\u5f0f\u9001\u51fa 5 \u4e2a\u4fdd\u6e29\u676f<\/li>\n<\/ul>\n<p><img decoding=\"async\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"http:\/\/4563.org\/wp-content\/uploads\/2021\/04\/20210419_607d36a5712db.jpg\" alt=\"\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848\" \/><\/p>\n<h2>\u5c42\u62bd\u5956 \u2014\u2014 \u963f\u91cc\u4e91\u89c6\u9891\u4e91\u5927\u8d5b\u9650\u91cf\u7535\u8111\u5305<\/h2>\n<ul>\n<li>\u8bc4\u8bba\u56de\u590d\u5185\u5bb9\u4e0d\u9650<\/li>\n<li>\u62bd 5 \u4e2a\u4eba\uff0c\u6bcf\u4eba\u9001\u4e00\u4e2a\u7535\u8111\u5305<\/li>\n<li>\u622a\u6b62\u65e5\u671f\u4e3a\uff1a2021.4.12 15:00<\/li>\n<li>\u4ece\u56de\u590d\u5c42\u4e2d\u968f\u673a\u62bd\u53d6<\/li>\n<li>\u4e2d\u5956\u7ed3\u679c\u4f1a\u4ee5\u9644\u8a00\u5f62\u5f0f\u516c\u5e03\u4e8e\u672c\u5e16\uff0c\u5e76 @ \u5404\u4f4d\u4e2d\u5956\u7528\u6237<\/li>\n<\/ul>\n<pre><code>function createRandom(num,from,to) {     var arr=[];      var json={};       while(arr.length&lt;num)     {         var ranNum=Math.round(Math.random()*(to-from))+from;         if(!json[ranNum])         {             json[ranNum]=1;             arr.push(ranNum);          }     }     return arr; } createRandom(5,0, \u56de\u590d\u5c42) \/\/ \u62bd\u5956 <\/code><\/pre>\n<p>\u6e90\u7801\u5f15\u81ea yedanbo\/createRandom().js<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" referrerpolicy=\"no-referrer\" rel=\"noreferrer\" src=\"http:\/\/4563.org\/wp-content\/uploads\/2021\/04\/20210419_607d36a8912fe.jpg\" alt=\"\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b\u65b9\u6848\" \/><\/p>\n<\/p><\/div>\n<div> <b>\u5927\u4f6c\u6709\u8a71\u8aaa<\/b> (<span>2<\/span>)        <\/div>\n<div> <\/div>\n<\/p><\/div>\n<\/p><\/div>\n<ul>\n<li data-pid=\"5745750\" data-uid=\"2\">\n<div>\n<div>\n<div> <span>\u8cc7\u6df1\u5927\u4f6c : homu <\/span>  <\/div>\n<div> <i title=\"\u5f15\u7528\"><\/i>  <span>          <\/span> <\/div>\n<\/p><\/div>\n<div>                                                             \u4e0a\u6b21\u5206\u6bcd\u4e86\uff0c\u518d\u6765\u8bd5\u8bd5                                                            <\/div>\n<\/p><\/div>\n<\/li>\n<li data-pid=\"5745751\" data-uid=\"2\">\n<div>\n<div>\n<div> <span>\u8cc7\u6df1\u5927\u4f6c : luanqiao <\/span>  <\/div>\n<div> <i title=\"\u5f15\u7528\"><\/i>  <span>          <\/span> <\/div>\n<\/p><\/div>\n<div>                                                             \u6211\u662f\u4e8c                                                            <\/div>\n<\/p><\/div>\n<\/li>\n<li>\n","protected":false},"excerpt":{"rendered":"<p>\u76d6\u62bd\u5956\uff5c 5 \u5206\u949f\u5b66\u4e60\u963f\u91cc\u7b97\u6cd5\u51a0\u519b&hellip;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[],"tags":[],"_links":{"self":[{"href":"http:\/\/4563.org\/index.php?rest_route=\/wp\/v2\/posts\/437375"}],"collection":[{"href":"http:\/\/4563.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/4563.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/4563.org\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/4563.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=437375"}],"version-history":[{"count":0,"href":"http:\/\/4563.org\/index.php?rest_route=\/wp\/v2\/posts\/437375\/revisions"}],"wp:attachment":[{"href":"http:\/\/4563.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=437375"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/4563.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=437375"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/4563.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=437375"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}