{"id":702,"date":"2024-01-27T19:28:59","date_gmt":"2024-01-27T11:28:59","guid":{"rendered":"https:\/\/scutvk.cn\/?p=702"},"modified":"2024-01-27T19:29:03","modified_gmt":"2024-01-27T11:29:03","slug":"%e9%98%85%e8%af%bb%e7%ac%94%e8%ae%b0tsns-towards-good-practices-for-deep-action-recognition","status":"publish","type":"post","link":"https:\/\/scutvk.cn\/?p=702","title":{"rendered":"[\u9605\u8bfb\u7b14\u8bb0]TSNs: Towards Good Practices for Deep Action Recognition"},"content":{"rendered":"

\u6587\u7ae0\u53d1\u8868\u4e8e2016 ECCV\uff0c\u4e3b\u8981\u63a2\u8ba8\u4e86\u6df1\u5ea6\u884c\u4e3a\u8bc6\u522b\u9886\u57df\u4e2d\u7684\u6700\u4f73\u5b9e\u8df5\u3002\u6587\u7ae0\u6307\u51fa\uff0c\u73b0\u6709\u57fa\u4e8eCNN\u7684\u5206\u7c7b\u65b9\u6cd5\u5728\u957f\u671f\u65f6\u95f4\u7ed3\u6784\u7684\u6355\u83b7\u4e0a\u4e0d\u5982\u624b\u5de5\u8bbe\u8ba1\u7684\u7279\u5f81\uff0c\u5e76\u56e0\u89c6\u9891\u6570\u636e\u96c6\u91cf\u4e0d\u8db3\u53ef\u80fd\u5bfc\u81f4\u8fc7\u62df\u5408\u3002\u4f5c\u4e3a\u89e3\u51b3\u65b9\u6848\uff0c\u6587\u7ae0\u63d0\u51fa\u4e86\u4e00\u4e2a\u9ad8\u6548\u6a21\u578b\uff0c\u4e13\u6ce8\u4e8e\u957f\u671f\u65f6\u95f4\u7ed3\u6784\u7684\u6355\u83b7\uff0c\u5e76\u8ba8\u8bba\u4e86\u5728\u6570\u636e\u6709\u9650\u7684\u60c5\u51b5\u4e0b\u5982\u4f55\u6709\u6548\u8bad\u7ec3\u6a21\u578b\uff0c\u5305\u62ec\u8de8\u6a21\u6001\u9884\u8bad\u7ec3\u3001\u89c4\u8303\u5316\u6280\u672f\u548c\u6570\u636e\u589e\u5f3a\uff0c\u540c\u65f6\u8fd8\u4ecb\u7ecd\u4e86\u7b2c\u4e00\u4e2a\u7aef\u5230\u7aef\u7684\u89c6\u9891\u65f6\u5e8f\u5efa\u6a21\u6a21\u578b\u3002<\/p>\n

1 Introduction<\/h2>\n

\u5728\u5206\u7c7b\u4e2d\uff0c\u73b0\u6709\u7684CNN based\u6027\u80fd\u8fd8\u6bd4\u4e0d\u8fc7\u624b\u5de5\u8bbe\u8ba1feature\u7684\u65b9\u6cd5\uff0c\u4f5c\u8005\u8ba4\u4e3a\u4e3b\u8981\u5f52\u7ed3\u4e8e2\u65b9\u9762\u7684\u539f\u56e0\uff1a<\/p>\n

    \n
  1. Long-Range temporal structure\u5f88\u91cd\u8981\uff0c\u4f46\u662f\u73b0\u5728\u7684\u4e3b\u8981\u8fd8\u5728\u5173\u6ce8Short-Term motions\u548cAppearance\u3002<\/li>\n
  2. \u73b0\u6709\u7684\u89c6\u9891\u6570\u636e\u96c6\u91cf\u4e0d\u591f\uff0c\u53ef\u80fd\u4f1a\u8fc7\u62df\u5408\u3002<\/li>\n<\/ol>\n

    2 Contributions<\/h2>\n
      \n
    1. \u63d0\u51fa\u4e00\u4e2a\u9ad8\u6548\u7684\uff0clong-range temporal structure \u6355\u83b7\u80fd\u529b\u7684\u6a21\u578b\u3002<\/li>\n
    2. \u600e\u4e48\u5728\u6709\u9650\u7684\u6570\u636e\u4e0b\u6709\u6548\u5730\u8bad\u7ec3\u6a21\u578b\uff08Cross Modality Pre-training\u3001Regularization Techniques\u3001Data Augmentation\uff09\u3002<\/li>\n
    3. \u7b2c\u4e00\u4e2aend-to-end\u7684video\u65f6\u5e8f\u5efa\u6a21\u6a21\u578b\u3002<\/li>\n<\/ol>\n

      3 Method<\/h2>\n

      \"image-20231208120653265\"<\/p>\n

        \n
      1. \u5c06video\u5747\u5300\u5730\u5206\u5272\u6210K\u54e5Segment<\/li>\n
      2. \u6bcf\u4e2asegment\u91cc\u9762\u968f\u673a\u62bd\u53d6\u4e00\u5e27RGB\u56fe\u7247\u548c a stack of consecutive optical flow\uff08\u5b9e\u9a8c\u4e2d\u8bbe\u7f6e\u4e3a\u8fde\u7eed\u76845\u5e27\uff09\u3002<\/li>\n
      3. \u7ecf\u8fc7BN-Inception\u8fdb\u884c\u7279\u5f81\u63d0\u53d6\u3002<\/li>\n
      4. RGB\u3001Optical\u4e24\u4e2aStream\u5404\u81ea\u5bf9probabilities\u8fdb\u884cavg\u7136\u540esoftmax\u3002<\/li>\n
      5. \u878d\u5408\u540e\u7684probabilities\u518d\u8fdb\u884cavg\u3002<\/li>\n<\/ol>\n

        4 \u5c11\u6570\u636e\u6709\u6548\u8bad\u7ec3\u6a21\u578b\uff0c\u51cf\u7f13\u8fc7\u62df\u5408\u7684\u65b9\u6cd5<\/h2>\n

        4.1 Cross Modality Pre-training<\/h3>\n

        \u5728ImageNet\u4e0a\u8bad\u7ec3\u7684\u53c2\u6570\uff0c\u7ecf\u8fc7\u5c06RGB 3 channels\u7684\u53c2\u6570avg\u540e\uff0c\u590d\u5236\u6210Optical Flow\u9700\u8981\u7684channels\u6765\u4f7f\u7528\u3002<\/p>\n

        4.2 Regularization Techniques<\/h3>\n

        \u4f5c\u8005\u63d0\u5230\uff0c\u666e\u901a\u7684BN Layer\uff0c\u5bf9\u6570\u636e\u8fdb\u884cNorm\u6709\u5229\u4e8e\u52a0\u5feb\u8bad\u7ec3\uff0c\u4f46\u662f\uff0c\u8fd9\u4e2a\u6a21\u6001\u7684\u7279\u5f81\u5206\u5e03\u5b66\u4e60\u597d\u600e\u4e48Norm\u540e\uff0c\u8fc1\u79fb\u5230\u53e6\u5916\u4e00\u4e2a\u6a21\u6001\u91cc\u9762\u7684\u8bdd\uff0c\u7279\u5f81\u7684\u5206\u5e03\u6539\u53d8\u4e86\uff0c\u539f\u6765Norm\u7684\u53c2\u6570\u5c06\u4e0d\u518d\u80fd\u5e2e\u52a9\u65b0\u6a21\u6001\u6570\u636e\u6709\u6548Norm\u6210\u539f\u6765\u7684\u5206\u5e03\u3002<\/p>\n

        \u6539\u8fdb\uff1a\u9884\u8bad\u7ec3\u597d\u540e\uff0c\u5c06\u7b2c\u4e00\u5c42BN Layer\u4e4b\u5916\u7684BN Layers\u5168\u90e8\u51bb\u4f4f\uff0c\u7b2c\u4e00\u5c42\u7684BN Layer\u53ef\u4ee5\u5728\u65b0\u6a21\u6001\u4e2d\u8fdb\u884c\u5fae\u8c03\u3002<\/p>\n

        4.3 Data Augmentation<\/h3>\n

        \u63d0\u51fa2\u4e2a\u65b0\u6a21\u6001\uff1a<\/p>\n

          \n
        1. RGB Difference\uff0c\u8ba1\u7b97\u524d\u540e\u4e24\u5e27RGB image\u7684\u5dee\u8ddd\u3002\u6709\u7528\uff0c\u4f46\u4e0d\u7a33\u5b9a\uff0c\u53ef\u4ee5\u4f5c\u4e3a\u6ca1Optical Flow\u65f6\u5019\u7684\u8865\u5145\uff0c\u8ba1\u7b97\u5feb\uff0c\u5355\u72ec\u4e0d\u5982RGB\u3002<\/li>\n
        2. Warped flow\uff0c\u8ba1\u7b97\u5149\u6d41\u7684\u6574\u4f53\u52a8\u5411\u4e4b\u540e\uff0c\u6574\u4e2a\u5149\u6d41\u51cf\u53bb\u6574\u4f53\u52a8\u5411\u7684\u5e73\u5747\u503c\uff0c\u51cf\u5c11\u662f\u89c6\u89d2\u79fb\u52a8\u5e26\u6765\u7684\u5f71\u54cd\u3002\u6709\u7528\uff0c\u5355\u72ec\u4e0d\u5982Optical Flow\u3002<\/li>\n<\/ol>\n

          \"image-20231208122632012\"
          \n\"image-20231208122653947\"<\/p>\n

          5 Experiments<\/h2>\n

          5.1 \u6d88\u878d<\/h3>\n

          \"image-20231208122937810\"<\/p>\n

          5.2 \u53ef\u89c6\u5316<\/h3>\n

          \u4ee5\u4e0b\u662f\u4f5c\u8005\u8fdb\u884c\u6a21\u578b\u53ef\u89c6\u5316\u7684\u5177\u4f53\u6b65\u9aa4\uff1a<\/p>\n

          \u9009\u62e9\u767d\u566a\u58f0\u56fe\u50cf\u4f5c\u4e3a\u8d77\u70b9\uff1aDeepDraw \u5de5\u5177\u4ece\u4e00\u4e2a\u53ea\u5305\u542b\u767d\u566a\u58f0\u7684\u56fe\u50cf\u5f00\u59cb\uff0c\u8fd9\u610f\u5473\u7740\u521d\u59cb\u56fe\u50cf\u662f\u968f\u673a\u7684\uff0c\u4e0d\u5305\u542b\u4efb\u4f55\u7279\u5b9a\u7684\u7ed3\u6784\u6216\u6a21\u5f0f\u3002<\/p>\n

          \u8fed\u4ee3\u68af\u5ea6\u4e0a\u5347\uff1a\u4f7f\u7528\u68af\u5ea6\u4e0a\u5347\u65b9\u6cd5\u6765\u8c03\u6574\u56fe\u50cf\u4e2d\u7684\u50cf\u7d20\u503c\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u662f\u901a\u8fc7\u8ba1\u7b97\u7f51\u7edc\u5bf9\u4e8e\u67d0\u4e00\u7c7b\u522b\uff08\u4f8b\u5982\u67d0\u4e2a\u7279\u5b9a\u52a8\u4f5c\uff09\u7684\u54cd\u5e94\u5e76\u6839\u636e\u8fd9\u4e2a\u54cd\u5e94\u7684\u68af\u5ea6\u6765\u66f4\u65b0\u56fe\u50cf\u6765\u5b9e\u73b0\u7684\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5c31\u662f\u4fee\u6539\u56fe\u50cf\u4ee5\u589e\u52a0\u7f51\u7edc\u5bf9\u4e8e\u7279\u5b9a\u7c7b\u522b\u7684\u8f93\u51fa\u5f97\u5206\u3002<\/p>\n

          \u91cd\u590d\u8fed\u4ee3\u76f4\u81f3\u6536\u655b\uff1a\u8fd9\u4e2a\u8fc7\u7a0b\u5728\u591a\u6b21\u8fed\u4ee3\u4e2d\u91cd\u590d\uff0c\u6bcf\u6b21\u8fed\u4ee3\u90fd\u6839\u636e\u7f51\u7edc\u7684\u53cd\u9988\u6765\u66f4\u65b0\u56fe\u50cf\u3002\u968f\u7740\u8fed\u4ee3\u7684\u8fdb\u884c\uff0c\u56fe\u50cf\u9010\u6e10\u6f14\u53d8\u4e3a\u66f4\u80fd\u6fc0\u6d3b\u7f51\u7edc\u7279\u5b9a\u90e8\u5206\u7684\u5f62\u5f0f\u3002<\/p>\n

          \u751f\u6210\u7279\u5b9a\u7c7b\u522b\u7684\u53ef\u89c6\u5316\uff1a\u6700\u7ec8\uff0c\u7ecf\u8fc7\u591a\u6b21\u8fed\u4ee3\u540e\uff0c\u5f97\u5230\u7684\u56fe\u50cf\u53ef\u4ee5\u88ab\u8ba4\u4e3a\u662f\u5bf9\u7f51\u7edc\u4e2d\u6240\u5b66\u4e60\u7279\u5f81\u7684\u4e00\u79cd\u53ef\u89c6\u5316\u3002\u8fd9\u4e9b\u56fe\u50cf\u901a\u5e38\u4f1a\u7a81\u51fa\u90a3\u4e9b\u5bf9\u4e8e\u7f51\u7edc\u8bc6\u522b\u7279\u5b9a\u7c7b\u522b\u6700\u91cd\u8981\u7684\u89c6\u89c9\u6a21\u5f0f\u3002<\/p>\n

          \u9002\u914d\u5149\u6d41\u6a21\u578b\uff1a\u539f\u59cb\u7684 DeepDraw \u5de5\u5177\u662f\u4e3a\u5904\u7406RGB\u6570\u636e\u8bbe\u8ba1\u7684\u3002\u4e3a\u4e86\u53ef\u89c6\u5316\u57fa\u4e8e\u5149\u6d41\u7684\u6a21\u578b\uff0c\u4f5c\u8005\u5bf9\u5de5\u5177\u8fdb\u884c\u4e86\u8c03\u6574\uff0c\u4f7f\u5176\u80fd\u591f\u4e0e\u4ed6\u4eec\u7684\u65f6\u95f4\u6d41\u5377\u79ef\u7f51\u7edc\u4e00\u8d77\u5de5\u4f5c\u3002<\/p>\n

          \u901a\u8fc7\u8fd9\u79cd\u65b9\u6cd5\uff0c\u4f5c\u8005\u80fd\u591f\u53ef\u89c6\u5316\u4e0d\u540c\u914d\u7f6e\u4e0b\uff08\u4f8b\u5982\uff0c\u6709\u65e0\u9884\u8bad\u7ec3\uff0c\u4f7f\u7528\u4e0d\u540c\u8f93\u5165\u6a21\u6001\u7b49\uff09\u7684\u7a7a\u95f4\u6d41\uff08\u57fa\u4e8eRGB\uff09\u548c\u65f6\u95f4\u6d41\uff08\u57fa\u4e8e\u5149\u6d41\uff09\u5377\u79ef\u7f51\u7edc\u7684\u5b66\u4e60\u7279\u5f81\u3002\u8fd9\u4e9b\u53ef\u89c6\u5316\u7ed3\u679c\u6709\u52a9\u4e8e\u66f4\u597d\u5730\u7406\u89e3\u7f51\u7edc\u662f\u5982\u4f55\u54cd\u5e94\u4e0d\u540c\u7c7b\u578b\u7684\u52a8\u4f5c\uff0c\u5e76\u63ed\u793a\u4e86\u7f51\u7edc\u5982\u4f55\u5728\u957f\u671f\u65f6\u95f4\u7ed3\u6784\u5efa\u6a21\u65b9\u9762\u53d6\u5f97\u8fdb\u5c55\u3002<\/p>\n

          \"image-20231208123021234\"<\/p>\n","protected":false},"excerpt":{"rendered":"

          \u300aTSNs: Towards Good Practices for Deep Action Recognition\u300b, \u53d1\u8868\u4e8e2016 ECCV\uff0c\u4e3b\u8981\u63a2\u8ba8\u4e86\u6df1\u5ea6\u884c\u4e3a\u8bc6\u522b\u9886\u57df\u4e2d\u7684\u6700\u4f73\u5b9e\u8df5\u3002\u6587\u7ae0\u6307\u51fa\uff0c\u73b0\u6709\u57fa\u4e8eCNN\u7684\u5206\u7c7b\u65b9\u6cd5\u5728\u957f\u671f\u65f6\u95f4\u7ed3\u6784\u7684\u6355\u83b7\u4e0a\u4e0d\u5982\u624b\u5de5\u8bbe\u8ba1\u7684\u7279\u5f81\uff0c\u5e76\u56e0\u89c6\u9891\u6570\u636e\u96c6\u91cf\u4e0d\u8db3\u53ef\u80fd\u5bfc\u81f4\u8fc7\u62df\u5408\u3002\u4f5c\u4e3a\u89e3\u51b3\u65b9\u6848\uff0c\u6587\u7ae0\u63d0\u51fa\u4e86\u4e00\u4e2a\u9ad8\u6548\u6a21\u578b\uff0c\u4e13\u6ce8\u4e8e\u957f\u671f\u65f6\u95f4\u7ed3\u6784\u7684\u6355\u83b7\uff0c\u5e76\u8ba8\u8bba\u4e86\u5728\u6570\u636e\u6709\u9650\u7684\u60c5\u51b5\u4e0b\u5982\u4f55\u6709\u6548\u8bad\u7ec3\u6a21\u578b\uff0c\u5305\u62ec\u8de8\u6a21\u6001\u9884\u8bad\u7ec3\u3001\u89c4\u8303\u5316\u6280\u672f\u548c\u6570\u636e\u589e\u5f3a\uff0c\u540c\u65f6\u8fd8\u4ecb\u7ecd\u4e86\u7b2c\u4e00\u4e2a\u7aef\u5230\u7aef\u7684\u89c6\u9891\u65f6\u5e8f\u5efa\u6a21\u6a21\u578b\u3002<\/p>\n","protected":false},"author":1,"featured_media":703,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"jetpack_publicize_message":"","jetpack_is_tweetstorm":false,"jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","enabled":false}}},"categories":[2,35,31,34,36],"tags":[],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"https:\/\/scutvk.cn\/wp-content\/uploads\/2024\/01\/PixPin_2024-01-27_19-25-33.png","_links":{"self":[{"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/posts\/702"}],"collection":[{"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=702"}],"version-history":[{"count":1,"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/posts\/702\/revisions"}],"predecessor-version":[{"id":704,"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/posts\/702\/revisions\/704"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=\/wp\/v2\/media\/703"}],"wp:attachment":[{"href":"https:\/\/scutvk.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=702"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=702"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/scutvk.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=702"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}