{"id":34326,"date":"2025-08-22T16:38:49","date_gmt":"2025-08-22T08:38:49","guid":{"rendered":"https:\/\/aicats.wiki\/?p=34326"},"modified":"2025-08-22T16:39:02","modified_gmt":"2025-08-22T08:39:02","slug":"detr%e6%98%af%e4%bb%80%e9%ba%bc%ef%bc%9f-2025%e5%b9%b4%e7%9b%ae%e6%a8%99%e5%81%b5%e6%b8%ac%e6%8a%80%e8%a1%93%e8%a7%a3%e6%9e%90%e8%88%87%e6%87%89%e7%94%a8%e6%83%85%e5%a2%83%e5%85%a8%e6%8c%87%e5%8d%97","status":"publish","type":"post","link":"https:\/\/aicats.wiki\/tw\/2025\/08\/22\/34326-html","title":{"rendered":"DETR\u662f\u4ec0\u9ebc\uff1f 2025\u5e74\u76ee\u6a19\u5075\u6e2c\u6280\u8853\u89e3\u6790\u8207\u61c9\u7528\u60c5\u5883\u5168\u6307\u5357"},"content":{"rendered":"

\u672c\u6587\u5c08\u696d\u89e3\u6790DETR\uff08Detection Transformer\uff09\u76ee\u6a19\u5075\u6e2c\u6846\u67b6<\/strong>\u7684\u6280\u8853\u539f\u7406<\/a>\u3001\u4e3b\u6d41\u7d50\u69cb\u82072025\u5e74\u524d\u6cbf\u8fed\u4ee3\uff0c\u4e26\u6bd4\u8f03\u50b3\u7d71\u5075\u6e2c\u5668\u512a\u52e2\uff0c\u6db5\u84cbDeformable DETR\u3001DINO\u3001RT-DETR\u7b49\u5275\u65b0\u6a21\u578b<\/strong>\uff0c\u8a73\u8ff0\u5176\u5728\u667a\u6167\u57ce\u5e02\u3001\u5de5\u696d\u6aa2\u6e2c\u3001\u91ab\u5b78\u5f71\u50cf\u7b49\u591a\u884c\u696d\u7684\u5be6\u969b\u61c9\u7528\u3002\u5167\u5bb9\u5305\u542b\u6bd4\u8f03\u8868\u683c\u3001\u696d\u754c\u6e05\u55ae\u8207\u5be6\u7528\u958b\u6e90\u5de5\u5177\u63a8\u85a6\uff0c\u5354\u52a9AI<\/a>\u5f9e\u696d\u4eba\u54e1\u548c\u5de5\u7a0b\u5e2b\u5feb\u901f\u638c\u63e1\u6700\u65b0\u76ee\u6a19\u5075\u6e2c\u6280\u8853\u52d5\u5411\u3002<\/p>\n\n\n\n

\"DETR\u662f\u4ec0\u9ebc\uff1f<\/figure>\n\n\n\n

\u96a8\u8457\u4eba\u5de5\u667a\u6167\u6280\u8853\u548c\u6df1\u5ea6\u5b78\u7fd2\u5728\u96fb\u8166\u8996\u89ba\u9818\u57df\u7684\u5feb\u901f\u767c\u5c55\uff0c\u76ee\u6a19\u5075\u6e2c<\/strong>\u4e0d\u65b7\u8fce\u4f86\u65b0\u7684\u6280\u8853\u7a81\u7834\u3002DETR\uff08Detection Transformer\uff09<\/strong>\u4f5c\u70baTransformer\u67b6\u69cb\u5728\u76ee\u6a19\u5075\u6e2c\u9818\u57df\u7684\u91cd\u5927\u5275\u65b0\uff0c\u81ea2020\u5e74\u7531Facebook AI Research\u63d0\u51fa\u4ee5\u4f86\uff0c\u5df2\u6210\u70ba\u5b78\u8853\u548c\u5de5\u696d\u754c\u7684\u7814\u7a76\u71b1\u9ede\u3002\u672c\u6587\u5c07\u4ee5\u5c08\u696d\u65b0\u805e\u5831\u5c0e\u7684\u65b9\u5f0f\uff0c\u6df1\u5165\u89e3\u6790DETR\u7684\u6280\u8853\u539f\u7406\u3001\u7d50\u69cb\u7d44\u4ef6\u30012025\u5e74\u4e3b\u6d41\u6280\u8853\u8fed\u4ee3\uff0c\u4ee5\u53ca\u5728\u591a\u884c\u696d\u4e2d\u7684\u61c9\u7528\u5834\u666f\uff0c\u9644\u8868\u683c\u3001\u6e05\u55ae\u548c\u5be6\u7528\u93c8\u63a5\uff0c\u52a9\u529b\u8b80\u8005\u5feb\u901f\u638c\u63e1\u6700\u65b0\u6aa2\u6e2c\u6280\u8853\u52d5\u614b\u3002<\/p>\n\n\n\n

\n\n\n\n

DETR\u6280\u8853\u539f\u7406\u89e3\u6790<\/h2>\n\n\n\n

DETR\u6982\u8981\u8207\u6280\u8853\u80cc\u666f<\/h3>\n\n\n\n

DETR\uff08Detection Transformer\uff09<\/strong>\u662f\u4e00\u7a2e\u7aef\u5230\u7aef\u7684\u76ee\u6a19\u5075\u6e2c\u6846\u67b6<\/strong>\uff0c\u9996\u6b21\u5be6\u73fe\u4e86\u300c\u7121\u9700\u4eba\u5de5\u8a2d\u8a08\u9328\u6846\uff08Anchor\uff09\u300d\u548c\u300c\u7121\u9700\u975e\u6975\u5927\u503c\u6291\u5236\uff08NMS\uff09\u300d\u7684\u6975\u7c21\u5075\u6e2c\u8def\u5f91\u3002\u50b3\u7d71\u76ee\u6a19\u5075\u6e2c\u65b9\u6cd5\u5982Faster R-CNN\u3001YOLO\u7b49\uff0c\u901a\u5e38\u4f9d\u8cf4\u8907\u96dc\u7684\u5f8c\u8655\u7406\u548c\u9328\u6846\u8a2d\u8a08\uff0cDETR\u5247\u5b8c\u5168\u63a1\u7528Transformer\u7684\u300c\u96c6\u5408\u9810\u6e2c\u300d\u7406\u5ff5\uff0c\u6975\u5927\u7c21\u5316\u7cfb\u7d71\u67b6\u69cb\u3002<\/p>\n\n\n\n

DETR\u6838\u5fc3\u7d44\u4ef6\u53ca\u5de5\u4f5c\u6d41\u7a0b<\/h3>\n\n\n\n

\u4e0b\u8868\u7c21\u660e\u6bd4\u8f03\u4e86\u50b3\u7d71\u76ee\u6a19\u5075\u6e2c\u5668\u8207DETR<\/strong>\u7684\u4e3b\u8981\u7279\u5fb5\uff1a<\/p>\n\n\n\n

\u7279\u9ede<\/th>\u50b3\u7d71\u6aa2\u6e2c\u5668\uff08Faster R-CNN\/YOLO\uff09<\/th>DETR<\/strong><\/th><\/tr>
Anchor\u8a2d\u8a08<\/td>\u9700\u624b\u52d5\u9810\u8a2d<\/td>\u7121\u9700Anchor<\/strong><\/td><\/tr>
NMS\u5f8c\u8655\u7406<\/td>\u5fc5\u9808<\/td>\u7121\u9700NMS<\/strong><\/td><\/tr>
\u5168\u57df\u4e0a\u4e0b\u6587\u8a0a\u606f<\/td>\u5c40\u90e8\u7279\u5fb5\u70ba\u4e3b\uff08CNN\uff09<\/td>\u5168\u5c40\u611f\u77e5\uff08\u81ea\u8a3b\u610f\u529b\uff09<\/strong><\/td><\/tr>
\u9810\u6e2c\u65b9\u5f0f<\/td>\u5169\u968e\u6bb5\/\u591a\u968e\u6bb5<\/td>\u4e00\u7d44\u96c6\u5408\u9810\u6e2c<\/strong><\/td><\/tr>
\u53ef\u64f4\u5c55\u6027<\/td>\u8f03\u5dee<\/td>\u9ad8\u5ea6\u53ef\u64f4\u5c55<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

DETR\u7684\u6280\u8853\u6d41\u7a0b\u5206\u56db\u5927\u6a21\u7d44\uff1a<\/strong><\/p>\n\n\n\n

    \n
  1. CNN\u7279\u5fb5\u63d0\u53d6backbone\uff08\u5982ResNet-50\uff09<\/strong><\/li>\n\n\n\n
  2. \u4f4d\u7f6e\u7de8\u78bc<\/strong>\uff1a\u5c07\u7a7a\u9593\u8cc7\u8a0a\u878d\u5165\u7279\u5fb5\u5e8f\u5217<\/li>\n\n\n\n
  3. Transformer\u7de8\u78bc\u5668-\u89e3\u78bc\u5668<\/strong>\uff1a\u5168\u57df\u7279\u5fb5\u5efa\u6a21\u8207Object Query\u76ee\u6a19\u8868\u5fb5\u5b78\u7fd2<\/li>\n\n\n\n
  4. \u8f38\u51fa\u982d<\/strong>\uff1a\u900f\u904e\u96c6\u5408\u9810\u6e2c\u76f4\u63a5\u8f38\u51fa\u908a\u6846\u548c\u985e\u5225<\/li>\n<\/ol>\n\n\n\n
    \"\u958b\u6e90\u5009\u5eab\"
    \u5716\uff0f\u958b\u6e90\u5009\u5eab<\/a><\/figcaption><\/figure>\n\n\n\n

    \u6280\u8853\u9aa8\u5e79\u5be6\u73fe\u53c3\u8003\uff08PyTorch\u6e90\u78bc\u53ef\u898b\u958b\u6e90\u5009\u5eab<\/a>\uff09\uff1a<\/p>\n\n\n\n\n

    features = backbone(image) proj_features = projection(features) + positional_encoding memory = transformer_encoder(proj_features) outputs = transformer_decoder(object_queries, memory) detection = prediction_head(\n<\/code><\/pre>\n\n\n\n\n
    DETR\u4e2d\u7684Transformer\u8207Object Query<\/h3>\n\n\n\n