{"id":12150,"date":"2023-01-31T11:38:45","date_gmt":"2023-01-31T03:38:45","guid":{"rendered":"http:\/\/139.9.1.231\/?p=12150"},"modified":"2023-01-31T14:12:08","modified_gmt":"2023-01-31T06:12:08","slug":"monovit","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2023\/01\/31\/monovit\/","title":{"rendered":"MonoViT—\u57fa\u4e8eViT\u7684\u81ea\u76d1\u7763\u5355\u76ee\u6df1\u5ea6\u4f30\u8ba1"},"content":{"rendered":"\n

Self-Supervised Monocular Depth Estimation witha Vision Transformer<\/a><\/strong><\/p>\n\n\n\n

paper\uff1a https:\/\/arxiv.org\/pdf\/2208.03543.pdf<\/a><\/em><\/strong><\/p>\n\n\n\n

\u57fa\u4e8e\u6df1\u5ea6\u5b66\u4e60\u5355\u76ee\u6df1\u5ea6\u4f30\u8ba1\u4efb\u52a1\u7b80\u4ecb<\/h3>\n\n\n\n

\u6df1\u5ea6\u4f30\u8ba1\u662f\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u7684\u4e00\u4e2a\u57fa\u7840\u6027\u95ee\u9898\uff0c\u5176\u53ef\u4ee5\u5e94\u7528\u5728\u673a\u5668\u4eba\u5bfc\u822a\u3001\u589e\u5f3a\u73b0\u5b9e\u3001\u4e09\u7ef4\u91cd\u5efa\u3001\u81ea\u52a8\u9a7e\u9a76\u7b49\u9886\u57df\u3002\u800c\u76ee\u524d\u5927\u90e8\u5206\u6df1\u5ea6\u4f30\u8ba1\u90fd\u662f\u57fa\u4e8e\u4e8c\u7ef4RGB\u56fe\u50cf\u5230RBG-D\u56fe\u50cf\u7684\u8f6c\u5316\u4f30\u8ba1\uff0c\u4e3b\u8981\u5305\u62ec\u4ece\u56fe\u50cf\u660e\u6697\u3001\u4e0d\u540c\u89c6\u89d2\u3001\u5149\u5ea6\u3001\u7eb9\u7406\u4fe1\u606f\u7b49\u83b7\u53d6\u573a\u666f\u6df1\u5ea6\u5f62\u72b6\u7684Shape from X\u65b9\u6cd5\uff0c\u8fd8\u6709\u7ed3\u5408SFM(Structure from motion)\u548cSLAM(Simultaneous Localization And Mapping)\u7b49\u65b9\u5f0f\u9884\u6d4b\u76f8\u673a\u4f4d\u59ff\u7684\u7b97\u6cd5\u3002\u5176\u4e2d\u867d\u7136\u6709\u5f88\u591a\u8bbe\u5907\u53ef\u4ee5\u76f4\u63a5\u83b7\u53d6\u6df1\u5ea6\uff0c\u4f46\u662f\u8bbe\u5907\u9020\u4ef7\u6602\u8d35\u3002\u4e5f\u53ef\u4ee5\u5229\u7528\u53cc\u76ee\u8fdb\u884c\u6df1\u5ea6\u4f30\u8ba1\uff0c\u4f46\u662f\u7531\u4e8e\u53cc\u76ee\u56fe\u50cf\u9700\u8981\u5229\u7528\u7acb\u4f53\u5339\u914d\u8fdb\u884c\u50cf\u7d20\u70b9\u5bf9\u5e94\u548c\u89c6\u5dee\u8ba1\u7b97\uff0c\u6240\u4ee5\u8ba1\u7b97\u590d\u6742\u5ea6\u4e5f\u8f83\u9ad8\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u4f4e\u7eb9\u7406\u573a\u666f\u7684\u5339\u914d\u6548\u679c\u4e0d\u597d\u3002\u800c\u5355\u76ee\u6df1\u5ea6\u4f30\u8ba1\u5219\u76f8\u5bf9\u6210\u672c\u66f4\u4f4e\uff0c\u66f4\u5bb9\u6613\u666e\u53ca\u3002<\/p>\n\n\n\n

\u90a3\u4e48\u5bf9\u4e8e\u5355\u76ee\u6df1\u5ea6\u4f30\u8ba1\uff0c\u987e\u540d\u601d\u4e49\uff0c\u5c31\u662f\u5229\u7528\u4e00\u5f20\u6216\u8005\u552f\u4e00\u89c6\u89d2\u4e0b\u7684RGB\u56fe\u50cf\uff0c\u4f30\u8ba1\u56fe\u50cf\u4e2d\u6bcf\u4e2a\u50cf\u7d20\u76f8\u5bf9\u62cd\u6444\u6e90\u7684\u8ddd\u79bb\u3002\u5bf9\u4e8e\u4eba\u773c\u6765\u8bf4\uff0c\u7531\u4e8e\u5b58\u5728\u5927\u91cf\u7684\u5148\u9a8c\u77e5\u8bc6\uff0c\u6240\u4ee5\u53ef\u4ee5\u4ece\u4e00\u53ea\u773c\u775b\u6240\u83b7\u53d6\u7684\u56fe\u50cf\u4fe1\u606f\u4e2d\u63d0\u53d6\u51fa\u5927\u91cf\u6df1\u5ea6\u4fe1\u606f\u3002\u90a3\u4e48\u5355\u76ee\u6df1\u5ea6\u4f30\u8ba1\u4e0d\u4ec5\u9700\u8981\u4ece\u4e8c\u7ef4\u56fe\u50cf\u4e2d\u5b66\u4f1a\u5ba2\u89c2\u7684\u6df1\u5ea6\u4fe1\u606f\uff0c\u800c\u4e14\u9700\u8981\u63d0\u53d6\u4e00\u4e9b\u7ecf\u9a8c\u4fe1\u606f\uff0c\u540e\u8005\u5219\u5bf9\u4e8e\u6570\u636e\u96c6\u4e2d\u76f8\u673a\u548c\u573a\u666f\u4f1a\u6bd4\u8f83\u654f\u611f\u3002<\/p>\n\n\n\n

\u6458\u8981\uff1a<\/h3>\n\n\n\n

\u81ea\u76d1\u7763\u5355\u773c\u6df1\u5ea6\u4f30\u8ba1\u662f\u4e00\u79cd\u6709\u5438\u5f15\u529b\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u5b83\u4e0d\u9700\u8981\u96be\u4ee5\u83b7\u53d6\u7684\u6df1\u5ea6\u6807\u7b7e\u6765\u8fdb\u884c\u8bad\u7ec3\u3002 \u5377\u79ef\u795e\u7ecf\u7f51\u7edc (CNN) \u6700\u8fd1\u5728\u8fd9\u9879\u4efb\u52a1\u4e2d\u53d6\u5f97\u4e86\u5de8\u5927\u6210\u529f\u3002 \u7136\u800c\uff0c\u5b83\u4eec\u6709\u9650\u7684\u63a5\u53d7\u57df\u9650\u5236\u4e86\u73b0\u6709\u7684\u7f51\u7edc\u67b6\u6784\u53ea\u80fd\u5728\u5c40\u90e8\u8fdb\u884c\u63a8\u7406\uff0c\u4ece\u800c\u524a\u5f31\u4e86\u81ea\u6211\u76d1\u7763\u8303\u5f0f\u7684\u6709\u6548\u6027\u3002 \u9274\u4e8e Vision Transformers (ViTs) \u6700\u8fd1\u53d6\u5f97\u7684\u6210\u529f\uff0c\u6211\u4eec\u63d0\u51fa\u4e86 MonoViT\uff0c\u8fd9\u662f\u4e00\u4e2a\u5168\u65b0\u7684\u6846\u67b6\uff0c\u7ed3\u5408\u4e86 ViT \u6a21\u578b\u652f\u6301\u7684\u5168\u5c40\u63a8\u7406\u548c\u81ea\u76d1\u7763\u5355\u76ee\u6df1\u5ea6\u4f30\u8ba1\u7684\u7075\u6d3b\u6027\u3002 \u901a\u8fc7\u5c06\u666e\u901a\u5377\u79ef\u4e0e Transformer \u5757\u76f8\u7ed3\u5408\uff0c\u6211\u4eec\u7684\u6a21\u578b\u53ef\u4ee5\u5728\u5c40\u90e8\u548c\u5168\u5c40\u8fdb\u884c\u63a8\u7406\uff0c\u4ee5\u66f4\u9ad8\u7684\u7ec6\u8282\u548c\u51c6\u786e\u6027\u4ea7\u751f\u6df1\u5ea6\u9884\u6d4b\uff0c\u4ece\u800c\u4f7f MonoViT \u5728\u5df2\u5efa\u7acb\u7684 KITTI \u6570\u636e\u96c6\u4e0a\u5b9e\u73b0sota\u7684\u6027\u80fd\u3002 \u6b64\u5916\uff0cMonoViT \u5728 Make3D \u548c Driving Stereo \u7b49\u5176\u4ed6\u6570\u636e\u96c6\u4e0a\u8bc1\u660e\u4e86\u5176\u5353\u8d8a\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n

\u4ecb\u7ecd\uff1a<\/h3>\n\n\n\n

Transformers (ViTs)\u6700\u8fd1\u8868\u73b0\u51fa\u6770\u51fa\u7684\u76ee\u6807\u68c0\u6d4b\u548c \u8bed\u4e49\u5206\u5272\u7b49\u4efb\u52a1\u7684\u7ed3\u679c\uff0c\u8fd9\u8981\u5f52\u529f\u4e8e\u5b83\u4eec\u80fd\u591f\u5efa\u7acb\u50cf\u7d20\u4e4b\u95f4\u7684\u957f\u8ddd\u79bb\u5173\u7cfb\uff0c\u56e0\u6b64\u662f\u5168\u5c40\u611f\u53d7\u91ce\u3002\u53e6\u5916\uff0c\u6709\u76f8\u5173\u5de5\u4f5c\u5c06VIT\u5e94\u7528\u4e8e\u6df1\u5ea6\u4f30\u8ba1\uff0c\u4f46\u4e0d\u662f\u91c7\u7528\u81ea\u76d1\u7763\u5355\u76ee\u6df1\u5ea6\u4f30\u8ba1\u3002 \u672c\u6587\u5f25\u8865\u4e86\u8fd9\u4e2a\u7f3a\u5931\u7684\u6b65\u9aa4\uff0c\u63d0\u51fa\u4e86MonoViT architecture\u3002\u5b83\u5728\u5176\u9aa8\u5e72\u7f51\u4e2d\u7ed3\u5408\u4e86\u5377\u79ef\u5c42\u548c\u6700\u5148\u8fdb\u7684 (SoTA) MPViT\u5757\u30101\u3011\u8fdb\u800c\u5bf9\u56fe\u7247\u4e2d\u7684\u5c40\u90e8\u4fe1\u606f\uff08objects\uff09\u548c\u5168\u5c40\u4fe1\u606f\uff08\u524d\u666f\u548c\u80cc\u666f\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u4ee5\u53ca\u7269\u4f53\u4e4b\u95f4\uff09\u8fdb\u884c\u5efa\u6a21\u3002 \u8be5\u7b56\u7565\u4f7f\u6211\u4eec\u80fd\u591f\u6d88\u9664\u7531 CNN \u7f16\u7801\u5668\u7684\u6709\u9650\u611f\u77e5\u57df\u5f15\u8d77\u7684\u74f6\u9888\uff0c\u4ea7\u751f\u81ea\u7136\u66f4\u7ec6\u7c92\u5ea6\u7684\u9884\u6d4b\u3002<\/p>\n\n\n\n

\u4f5c\u8005\u5728KITTI dataset\u8fdb\u884c\u5b9e\u9a8c\uff0c\u8868\u73b0\u4f18\u4e8e\u5176\u4ed6sota\u6a21\u578b\uff0c\u8fd8\u5206\u6790\u4e86\u6a21\u578b\u6cdb\u5316\u80fd\u529b\u8de8\u4e0d\u540c\u7684\u6570\u636e\u96c6\uff0c\u5c06 MonoViT \u4e0e\u5b83\u5728 Make3D \u548c Driving-Stereo datasets\u8fdb\u884c\u6bd4\u8f83\uff0c\u4e5f\u7a81\u51fa\u663e\u793a\u4e86 MonoViT \u7684\u5353\u8d8a\u6cdb\u5316\u80fd\u529b<\/p>\n\n\n\n

\u6a21\u578b\u67b6\u6784\uff1a<\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n

Deep Network\uff1a<\/h4>\n\n\n\n

Joint CNN & Transformer Layer used in depthencoder\uff1a<\/strong><\/p>\n\n\n\n

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

PoseNet\uff1a<\/h4>\n\n\n\n

PoseNet \u503e\u5411\u4e8e\u7b80\u5355\u800c\u6709\u6548\u7684\u5b9e\u73b0\u3002 \u5177\u4f53\u6765\u8bf4\uff0c PoseNet \u4f7f\u7528 ResNet18\u30102\u3011\u7684\u8f7b\u91cf\u7ea7\u7ed3\u6784\u3002 \u63a5\u6536\u76f8\u90bb\u56fe\u50cf [I, I\u2020] \u4f5c\u4e3a\u8f93\u5165\uff0c\u8f93\u51fa\u89c6\u9891\u5e8f\u5217\u76f8\u90bb\u5e27\u4e4b\u95f4\u7684 6 DoF \u76f8\u5bf9\u4f4d\u59ff T\u3002\u8fd9\u4e2a\u7f51\u7edc\u7528\u4e8e\u6700\u7ec8\u8f85\u52a9\u8ba1\u7b97loss\uff0c\u63d0\u4f9b\u76d1\u7763\u4fe1\u606f\u3002<\/p>\n\n\n\n

Loss\u635f\u5931\u51fd\u6570<\/h3>\n\n\n\n

View reconstruction loss\uff1a<\/p>\n\n\n\n

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

Smoothness loss. As in previous works\uff0c the edge-aware smoothness loss is used to improve the inverse depth map d:<\/p>\n\n\n\n

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

\u30101\u3011MPViT: Multi-Path Vision Transformer for Dense Prediction<\/a><\/p>\n\n\n\n

\u30102\u3011Deep residual learning for image recognition<\/p>\n","protected":false},"excerpt":{"rendered":"

Self-Supervised Monocular Depth Estimation witha Vision … \u7ee7\u7eed\u9605\u8bfbMonoViT—\u57fa\u4e8eViT\u7684\u81ea\u76d1\u7763\u5355\u76ee\u6df1\u5ea6\u4f30\u8ba1<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,35,36],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12150"}],"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=12150"}],"version-history":[{"count":111,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12150\/revisions"}],"predecessor-version":[{"id":12412,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12150\/revisions\/12412"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=12150"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=12150"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=12150"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}