\u6458\u8981：\u672c\u6587\u7cfb\u7edf\u68b3\u7406\u5362\u83c1\u535a\u58eb\u300a\u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b\u300b\u7b2c\u4e94\u6a21\u5757\u300aGLIP \u6a21\u578b\u8be6\u89e3\u300b（\u7b2c 16\u201319 \u96c6）\u7684\u6838\u5fc3\u5185\u5bb9\u3002GLIP \u901a\u8fc7\u4e09\u5927\u521b\u65b0\u5b9e\u73b0\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b：\u4e00\u662f\u67b6\u6784\u4e09\u4ef6\u5957（DyHead + BERT + \u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408），\u628a\u76ee\u6807\u68c0\u6d4b\u91cd\u6784\u4e3a\u77ed\u8bed\u5b9a\u4f4d（phrase grounding）；\u4e8c\u662f\u8054\u5408\u635f\u5931\u51fd\u6570（\u8de8\u6a21\u6001\u5bf9\u9f50\u635f\u5931 + \u5b9a\u4f4d\u635f\u5931），\u8ba9\u5206\u7c7b\u4ece\u56fa\u5b9a\u7c7b\u522b\u8868\u53d8\u4e3a\u5728\u63d0\u793a\u6587\u672c\u4e2d\u627e\u8bcd；\u4e09\u662f 27M \u9884\u8bad\u7ec3\u6570\u636e\u914d\u65b9（3M \u4eba\u5de5\u6807\u6ce8 + 24M \u56fe\u6587\u5bf9\u81ea\u8bad\u7ec3\u4f2a\u6846）\u3002\u6587\u7ae0\u4ece DETR\u3001YOLOS \u7684\u68c0\u6d4b Transformer \u5316\u8bb2\u8d77，\u5e76\u4e0e YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4，\u63ed\u793a\u4ece\u5c01\u95ed\u7c7b\u522b\u5230\u5f00\u653e\u8bcd\u6c47\u7684\u8303\u5f0f\u8dc3\u8fc1，\u6700\u540e\u4ecb\u7ecd GLIP \u4e0e Grounded-SAM \u7ec4\u5408\u7684\u843d\u5730\u6d41\u7a0b，\u5e2e\u52a9\u8bfb\u8005\u638c\u63e1 DETR \u539f\u7406\u5e76\u5feb\u901f\u4e0a\u624b\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218\u3002<\/p>\n
\u2705 \u7b2c\u4e94\u6a21\u5757\u300aGLIP \u6a21\u578b\u8be6\u89e3\u300b（\u7b2c 16\u201319 \u96c6）<\/p>\n
\u674e\u6c90\u6df1\u5ea6\u5b66\u4e60191\u96c6\u8bfe\u7a0b\u5168\u89e3\u6790：\u6a21\u5757\u62c6\u89e3\u3001\u5b66\u4e60\u8def\u5f84-CSDN\u535a\u5ba2<\/h3>\n\u5434\u6069\u8fbe\u300a\u9762\u5411\u5f00\u53d1\u8005\u7684\u63d0\u793a\u8bcd\u5de5\u7a0b\u300b-CSDN\u535a\u5ba2<\/h3>\n\u5434\u6069\u8fbe MCP \u6559\u7a0b（Model Context Protocol）（\u4e00）-CSDN\u535a\u5ba2<\/h3>\n
\u5434\u6069\u8fbe MCP \u6559\u7a0b（Model Context Protocol）（\u4e00）-CSDN\u535a\u5ba2<\/h3>\n
\u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b\u5b66\u4e60\u7b14\u8bb0 \u2014 ViT \u00b7 CLIP \u00b7 SAM \u00b7 GLIP \u00b7 Stable Diffusion<\/p>\n
\n
GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b：\u4ece YOLOS\/DETR \u5230\u77ed\u8bed\u5b9a\u4f4d<\/h3>\n
\u4ece YOLOS \/ DETR \u5230 GLIP\u2014\u2014\u628a\u76ee\u6807\u68c0\u6d4b\u91cd\u6784\u4e3a\u77ed\u8bed\u5b9a\u4f4d（\u68c0\u6d4b\u5373 grounding），\u7528\u8bed\u8a00\u5f00\u653e\u68c0\u6d4b\u5668\u7684\u8bcd\u6c47\u8868\u3002<\/p>\n
\u53ef\u89c6\u5316\u9875\u9762<\/h3>\n![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261001085022-6abe1ece5dac9.png\")
GLIP \u5f00\u653e\u8bcd\u6c47\u76ee\u6807\u68c0\u6d4b\u7cbe\u8bb2（\u542b YOLOS\/DETR \u539f\u7406） | \u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b \u7b2c 16\u201319 \u96c6<\/h3>\n\n- \u6587\u4ef6：artifacts\/glip-model-guide.html（\u6697\u8272\u8bbe\u8ba1，\u542b\u5206\u96c6\u5bfc\u822a\u3001GLIP \u67b6\u6784\u6d41\u7a0b\u56fe\u3001\u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408\u793a\u610f\u3001\u635f\u5931\u51fd\u6570\u516c\u5f0f\u5757\u3001\u56db\u6a21\u578b\u5bf9\u6bd4\u8868\u3001\u8bfe\u7a0b\u8109\u7edc\u65f6\u95f4\u7ebf，\u4ee3\u7801\u7ea7\u8d28\u68c0 PASS）<\/li>\n<\/ul>\n
\u5206\u96c6\u7ed3\u6784\u5b9e\u8bc1（\u6765\u6e90：\u8bfe\u7a0b\u9009\u96c6 \u94fe\u63a5）：<\/p>\n
\n\u96c6\u6570\u6807\u9898\u65f6\u957f<\/tr>\n \n\n\u7b2c 16 \u96c6<\/td>\n GLIP \u6a21\u578b\u8be6\u89e3(1) YOLOS \u548c DETR \u6a21\u578b\u539f\u7406\u89e3\u6790<\/td>\n 11.85 \u5206\u949f<\/td>\n<\/tr>\n \n\u7b2c 17 \u96c6<\/td>\n GLIP \u6a21\u578b\u8be6\u89e3(2) GLIP \u6a21\u578b\u539f\u7406\u89e3\u6790<\/td>\n 11.53 \u5206\u949f<\/td>\n<\/tr>\n \n\u7b2c 18 \u96c6<\/td>\n GLIP \u6a21\u578b\u8be6\u89e3(3) GLIP \u635f\u5931\u51fd\u6570<\/td>\n 12.43 \u5206\u949f<\/td>\n<\/tr>\n \n\u7b2c 19 \u96c6<\/td>\n GLIP \u6a21\u578b\u8be6\u89e3(4) GLIP \u4e0e YOLOS\/DETR\/YOLOv \u5bf9\u6bd4<\/td>\n 9.03 \u5206\u949f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n\u8be6\u7ec6\u603b\u7ed3（\u7ea6 5200 \u5b57）<\/h3>\n\u4e00\u3001\u6a21\u5757\u5b9a\u4f4d：\u4ece"\u5206\u5272\u4e00\u5207"\u5230"\u68c0\u6d4b\u4e00\u5207"<\/h4>\n
\u672c\u6a21\u5757\u4e3a\u5362\u83c1\u535a\u58eb\u300a\u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b\u300b\u7b2c\u4e94\u6a21\u5757，\u5171 4 \u96c6，\u603b\u65f6\u957f\u7ea6 44.8 \u5206\u949f\u3002\u4e0a\u4e00\u6a21\u5757 SAM \u89e3\u51b3\u4e86"\u7c7b\u522b\u65e0\u5173\u7684\u5206\u5272"，\u4f46\u7559\u4e0b\u4e00\u4e2a\u660e\u786e\u7f3a\u53e3：SAM \u53ea\u753b\u8f6e\u5ed3\u3001\u4e0d\u8ba4\u7c7b\u522b\u3002\u672c\u6a21\u5757\u8bb2\u89e3\u7684 GLIP（Grounded Language-Image Pre-training，\u5fae\u8f6f\u7814\u7a76\u9662，CVPR 2022）\u6070\u597d\u8865\u4e0a\u8fd9\u5757\u62fc\u56fe\u2014\u2014\u7528\u81ea\u7136\u8bed\u8a00\u5b9a\u4e49"\u8981\u68c0\u6d4b\u4ec0\u4e48"，\u628a\u5f00\u653e\u8bcd\u6c47\u80fd\u529b\u5e26\u5230\u5bf9\u8c61\u7ea7\u5b9a\u4f4d\u4efb\u52a1\u3002<\/p>\n
\u8bfe\u7a0b\u7f16\u6392\u9887\u5177\u5320\u5fc3：\u7b2c 16 \u96c6\u5e76\u4e0d\u76f4\u63a5\u8bb2 GLIP，\u800c\u662f\u5148\u94fa\u57ab\u4e24\u4ee3\u68c0\u6d4b Transformer（DETR\u3001YOLOS），\u56e0\u4e3a GLIP \u7684\u68c0\u6d4b\u9aa8\u5e72\u6b63\u662f\u5efa\u7acb\u5728\u8fd9\u6761\u6280\u672f\u7ebf\u4e0a\u3002\u7406\u89e3\u4e86"\u68c0\u6d4b\u5982\u4f55 Transformer \u5316"，\u624d\u80fd\u7406\u89e3 GLIP \u5728\u5176\u4e0a\u505a\u7684\u8303\u5f0f\u9769\u65b0\u3002<\/p>\n
\u4e8c\u3001\u7b2c 16 \u96c6：YOLOS \u4e0e DETR\u2014\u2014\u68c0\u6d4b\u7684 Transformer \u5316<\/h4>\n
DETR（2020，Facebook AI）：\u9996\u4e2a\u7aef\u5230\u7aef\u76ee\u6807\u68c0\u6d4b Transformer\u3002\u6d41\u7a0b\u4e3a：CNN \u9aa8\u5e72\u63d0\u53d6\u56fe\u50cf\u7279\u5f81 \u2192 Transformer \u7f16\u7801\u5668\u505a\u5168\u5c40\u5efa\u6a21 \u2192 \u89e3\u7801\u5668\u7528\u4e00\u7ec4\u56fa\u5b9a\u6570\u91cf（\u5982 100 \u4e2a）\u7684 object query \u5e76\u884c"\u8be2\u95ee"\u56fe\u50cf\u4e2d\u6bcf\u4e2a\u76ee\u6807\u7684\u4f4d\u7f6e \u2192 \u8f93\u51fa\u8fb9\u754c\u6846\u4e0e\u7c7b\u522b\u3002\u5b83\u6709\u4e24\u5927\u9769\u547d\u6027\u8bbe\u8ba1：<\/p>\n
- \u629b\u5f03\u951a\u6846（anchor）\u3002\u4f20\u7edf\u68c0\u6d4b\u5668（Faster R-CNN\u3001YOLO）\u4f9d\u8d56\u4eba\u5de5\u8bbe\u8ba1\u7684\u5148\u9a8c\u951a\u6846，\u5c3a\u5bf8\u3001\u6bd4\u4f8b\u3001\u6570\u91cf\u90fd\u662f\u8d85\u53c2\u6570；DETR \u7684 object query \u662f\u4e00\u7ec4\u53ef\u5b66\u4e60\u7684\u4f4d\u7f6e\u67e5\u8be2\u5411\u91cf，\u6a21\u578b\u81ea\u5df1\u5b66\u4f1a"\u53bb\u54ea\u91cc\u770b"，\u4e0d\u518d\u4f9d\u8d56\u4eba\u5de5\u5148\u9a8c\u3002<\/li>\n
- \u629b\u5f03 NMS（\u975e\u6781\u5927\u503c\u6291\u5236）\u3002\u4f20\u7edf\u68c0\u6d4b\u5668\u4f1a\u5bf9\u540c\u4e00\u76ee\u6807\u8f93\u51fa\u591a\u4e2a\u91cd\u53e0\u6846，\u518d\u7528 NMS \u8d2a\u5fc3\u53bb\u91cd；DETR \u7528**\u5308\u7259\u5229\u5339\u914d（\u4e8c\u5206\u56fe\u6700\u4f18\u5339\u914d）**\u5b9e\u73b0\u9884\u6d4b\u4e0e\u771f\u503c\u7684\u4e00\u5bf9\u4e00\u5206\u914d\u2014\u2014\u6bcf\u4e2a\u771f\u503c\u6846\u53ea\u7531\u4e00\u4e2a\u9884\u6d4b\u8d1f\u8d23，\u4ece\u673a\u5236\u6839\u4e0a\u6d88\u9664\u4e86\u91cd\u590d\u6846，\u5b9e\u73b0\u4e86\u771f\u6b63\u7684\u7aef\u5230\u7aef\u3002<\/li>\n
\u4ee3\u4ef7\u4e5f\u5f88\u660e\u663e：\u6536\u655b\u6162（\u9700\u6570\u767e\u4e2a epoch）\u3001\u5c0f\u76ee\u6807\u68c0\u6d4b\u504f\u5f31\u3002\u8fd9\u4e9b\u95ee\u9898\u50ac\u751f\u4e86\u540e\u7eed\u5927\u91cf\u6539\u8fdb\u5de5\u4f5c（Deformable DETR \u7b49），\u4f46"\u7aef\u5230\u7aef + \u96c6\u5408\u9884\u6d4b"\u7684\u8303\u5f0f\u7531\u6b64\u786e\u7acb\u3002<\/p>\n
YOLOS（2021）：\u4e00\u4e2a"\u6781\u7b80\u5b9e\u9a8c"\u2014\u2014\u62ff\u4e3a\u5206\u7c7b\u9884\u8bad\u7ec3\u7684\u88f8 ViT，\u51e0\u4e4e\u4e0d\u6539\u67b6\u6784，\u53ea\u8ffd\u52a0\u5c11\u91cf\u53ef\u5b66\u4e60\u7684 [DET] token，\u5c31\u76f4\u63a5\u505a\u76ee\u6807\u68c0\u6d4b，\u4e14\u6027\u80fd\u53ef\u89c2\u3002\u5b83\u7684\u610f\u4e49\u4e0d\u5728\u5237\u699c，\u800c\u5728\u8bc1\u660e：\u68c0\u6d4b\u80fd\u529b\u53ef\u4ee5\u4ece\u7eaf\u7cb9\u7684\u5e8f\u5217\u5230\u5e8f\u5217\u5efa\u6a21\u4e2d"\u6d8c\u73b0"，\u4e0d\u9700\u8981 CNN \u7684\u5f52\u7eb3\u504f\u7f6e，\u4e5f\u4e0d\u9700\u8981\u68c0\u6d4b\u4e13\u7528\u67b6\u6784\u3002\u8fd9\u4e0e\u8bfe\u7a0b\u7b2c\u4e00\u6a21\u5757 ViT"Transformer \u53ef\u5b8c\u5168\u66ff\u4ee3 CNN \u505a\u89c6\u89c9"\u7684\u7ed3\u8bba\u4e00\u8109\u76f8\u627f，\u628a"Transformer \u901a\u7528\u6027"\u4ece\u5206\u7c7b\u63a8\u8fdb\u5230\u4e86\u5b9a\u4f4d\u4efb\u52a1\u3002<\/p>\n
\u8fd9\u4e00\u96c6\u5408\u8d77\u6765\u56de\u7b54\u7684\u95ee\u9898\u662f：\u68c0\u6d4b\u5668\u5df2\u7ecf Transformer \u5316\u4e86，\u4f46\u5b83\u4ecd\u662f"\u5c01\u95ed\u4e16\u754c"\u7684\u2014\u2014\u7c7b\u522b\u8868\u56fa\u5b9a（\u5982 COCO 80 \u7c7b），\u6362\u4efb\u52a1\u5c31\u8981\u6539\u5206\u7c7b\u5934\u3001\u91cd\u65b0\u8bad\u7ec3\u3002\u8fd9\u5c31\u662f GLIP \u8981\u89e3\u51b3\u7684\u6839\u672c\u95ee\u9898\u3002<\/p>\n
\u4e09\u3001\u7b2c 17 \u96c6：GLIP \u539f\u7406\u2014\u2014\u68c0\u6d4b\u5373\u77ed\u8bed\u5b9a\u4f4d<\/h4>\n
GLIP \u7684\u6838\u5fc3\u601d\u60f3\u7528\u4e00\u53e5\u8bdd\u6982\u62ec：\u628a\u76ee\u6807\u68c0\u6d4b\u91cd\u6784\u4e3a\u77ed\u8bed\u5b9a\u4f4d（phrase grounding）\u4efb\u52a1\u3002<\/p>\n
\n- \u8f93\u5165：\u56fe\u50cf + \u4e00\u6bb5\u6587\u672c\u63d0\u793a，\u5982 cat . dog . person .（\u5404\u7c7b\u522b\u7528\u53e5\u53f7\u5206\u9694\u7684\u63d0\u793a\u53e5）<\/li>\n
- \u8f93\u51fa：\u56fe\u4e2d\u6bcf\u4e2a\u4e0e\u63d0\u793a\u4e2d\u77ed\u8bed\u8bed\u4e49\u5bf9\u5e94\u7684\u76ee\u6807\u6846<\/li>\n<\/ul>\n
\u5728 GLIP \u6846\u67b6\u4e0b，\u201c\u68c0\u6d4b"\u88ab\u7edf\u4e00\u8fdb grounding \u7684\u8bed\u4e49：\u4f20\u7edf\u68c0\u6d4b\u5668\u8f93\u51fa"\u8fd9\u662f\u7c7b\u522b\u7f16\u53f7 #15\u201d，GLIP \u8f93\u51fa"\u8fd9\u4e2a\u533a\u57df\u4e0e\u63d0\u793a\u4e2d\u7684\u8bcd \u2018cat\u2019 \u5bf9\u9f50"\u3002\u5206\u7c7b\u5934\u88ab\u66ff\u6362\u4e3a\u89c6\u89c9\u533a\u57df\u7279\u5f81\u4e0e\u6587\u672c\u8bcd\u7279\u5f81\u7684\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u2014\u2014\u68c0\u6d4b\u7684"\u7c7b\u522b\u8868"\u4ece\u6b64\u53d8\u6210\u4efb\u610f\u81ea\u7136\u8bed\u8a00，\u8bcd\u6c47\u8868\u5f7b\u5e95\u5f00\u653e：\u9762\u5bf9\u8bad\u7ec3\u96c6\u4e2d\u4ece\u672a\u51fa\u73b0\u7684\u65b0\u7c7b\u522b（\u5982"\u6234\u592a\u9633\u955c\u7684\u67f4\u72ac"），\u53ea\u9700\u6539\u63d0\u793a\u8bcd，\u96f6\u6837\u672c\u5373\u53ef\u68c0\u6d4b\u3002<\/p>\n
\u67b6\u6784\u4e09\u4ef6\u5957：<\/p>\n
- \u89c6\u89c9\u7f16\u7801\u5668：DyHead\u3002\u52a8\u6001\u5934\u7ed3\u6784，\u5728\u5c3a\u5ea6\u3001\u7a7a\u95f4\u3001\u4efb\u52a1\u4e09\u4e2a\u7ef4\u5ea6\u4e0a\u505a\u6ce8\u610f\u529b，\u4ea7\u51fa\u533a\u57df\u7ea7\u89c6\u89c9\u7279\u5f81\u3002<\/li>\n
- \u6587\u672c\u7f16\u7801\u5668：BERT（\u6216 CLIP \u6587\u672c\u7f16\u7801\u5668）\u3002\u628a\u63d0\u793a\u53e5\u7f16\u7801\u4e3a\u8bcd\u7ea7\u7279\u5f81\u5e8f\u5217\u3002<\/li>\n
- \u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408（language-aware deep fusion）\u3002\u8fd9\u662f GLIP \u533a\u522b\u4e8e CLIP \u7684\u5173\u952e\u3002CLIP \u53ea\u5728\u6700\u540e\u5bf9\u5168\u5c40\u7279\u5f81\u505a\u70b9\u79ef（late fusion，\u201c\u5404\u5b66\u5404\u7684\u3001\u6700\u540e\u5bf9\u7b54\u6848\u201d），\u662f\u56fe\u50cf\u7ea7\u7684\u6d45\u5bf9\u9f50；GLIP \u5219\u5728\u4e24\u4e2a\u7f16\u7801\u5668\u7684\u591a\u4e2a\u4e2d\u95f4\u5c42\u63d2\u5165\u8de8\u6a21\u6001\u4ea4\u53c9\u6ce8\u610f\u529b\u2014\u2014\u89c6\u89c9\u5c42\u7528\u6587\u672c\u7279\u5f81\u66f4\u65b0\u81ea\u5df1（"\u7ea2\u8272"\u8fd9\u4e2a\u8bcd\u4f1a\u5b9e\u65f6\u589e\u5f3a\u7ea2\u8272\u533a\u57df\u7684\u7279\u5f81\u6743\u91cd），\u6587\u672c\u5c42\u4e5f\u7528\u89c6\u89c9\u7279\u5f81\u56de\u5199，\u53cc\u5411\u9010\u5c42\u4ea4\u6362\u4fe1\u606f，\u5b9e\u73b0\u5bf9\u8c61\u7ea7\u7684\u6df1\u5ea6\u8bed\u4e49\u8026\u5408\u3002<\/li>\n
\u9884\u8bad\u7ec3\u6570\u636e\u914d\u65b9（27M）：<\/p>\n
\n- 3M \u4eba\u5de5\u6807\u6ce8：Flickr30K Entities\u3001Objects365 \u7b49\u68c0\u6d4b\/\u5b9a\u4f4d\u6570\u636e，\u6846\u4e0e\u77ed\u8bed\u5929\u7136\u5bf9\u5e94，\u8d28\u91cf\u9ad8\u4f46\u89c4\u6a21\u6709\u9650；<\/li>\n
- 24M \u7f51\u7edc\u56fe\u6587\u5bf9：\u6ca1\u6709\u6846\u6807\u6ce8\u3002GLIP \u7528**\u81ea\u8bad\u7ec3（self-training）**\u65b9\u5f0f\u2014\u2014\u5148\u8bad\u51fa\u7684\u6a21\u578b\u7ed9\u56fe\u6587\u5bf9 Caption \u4e2d\u7684\u540d\u8bcd\u77ed\u8bed"\u6253\u4f2a\u6846"，\u8fc7\u6ee4\u4f4e\u7f6e\u4fe1\u5ea6\u7ed3\u679c\u540e\u56de\u7089\u8bad\u7ec3\u3002\u8fd9\u4e0e CLIP \u5229\u7528 4 \u4ebf\u56fe\u6587\u5bf9\u7684\u601d\u8def\u540c\u6e90，\u4f46\u628a\u5bf9\u9f50\u7c92\u5ea6\u4ece\u56fe\u50cf\u7ea7\u4e0b\u6c89\u5230\u4e86\u5bf9\u8c61\u7ea7\u3002<\/li>\n<\/ul>\n
\u6548\u679c：GLIP-L \u5728\u96f6\u6837\u672c COCO \u68c0\u6d4b\u4e0a\u8fbe\u5230 49.8 AP，\u8d85\u8fc7\u6709\u76d1\u7763\u8bad\u7ec3\u7684\u5f3a\u57fa\u7ebf；\u5728 13 \u4e2a\u975e\u5e38\u89c4\u76ee\u6807\u6570\u636e\u96c6\u4e0a\u5e73\u5747\u8d85\u8d8a\u6709\u76d1\u7763\u65b9\u6cd5，\u8bc1\u660e\u4e86\u5f00\u653e\u8bcd\u6c47\u8def\u7ebf\u7684\u5b9e\u7528\u4ef7\u503c\u3002<\/p>\n
\u56db\u3001\u7b2c 18 \u96c6：GLIP \u635f\u5931\u51fd\u6570<\/h4>\n
GLIP \u7684\u8bad\u7ec3\u76ee\u6807\u662f"\u5206\u7c7b\u4e0e\u5b9a\u4f4d\u7edf\u4e00"\u7684\u8054\u5408\u635f\u5931，\u7531\u4e24\u90e8\u5206\u7ec4\u6210：<\/p>\n
\u2460 \u8de8\u6a21\u6001\u5bf9\u9f50\u635f\u5931（classification as grounding）：\u5148\u6784\u5efa"\u533a\u57df \u00d7 \u8bcd"\u76f8\u4f3c\u5ea6\u77e9\u9635 S = O\u00b7P\u1d40（\u533a\u57df\u7279\u5f81 O \u4e0e\u8bcd\u7279\u5f81 P \u7684\u70b9\u79ef），\u7528\u5b83\u66ff\u4ee3\u4f20\u7edf\u7684\u56fa\u5b9a\u7c7b\u522b softmax \u5206\u7c7b\u5934\u3002\u5308\u7259\u5229\u5339\u914d\u786e\u5b9a\u9884\u6d4b\u6846\u4e0e\u771f\u503c\u77ed\u8bed\u7684\u914d\u5bf9\u540e，\u5728\u77e9\u9635\u4e0a\u505a token \u7ea7\u5bf9\u9f50：\u6b63\u6837\u672c\u5bf9（\u6846\u4e0e\u5bf9\u5e94\u77ed\u8bed\u7684\u8bcd）\u76f8\u4f3c\u5ea6\u63a8\u9ad8，\u8d1f\u6837\u672c\u5bf9\u63a8\u4f4e\u2014\u2014\u672c\u8d28\u662f\u4e00\u4e2a\u5bf9\u8c61\u7ea7\u7684\u5bf9\u6bd4\u5206\u7c7b\u635f\u5931（\u7126\u70b9\u52a0\u6743\u4ee5\u5904\u7406\u6b63\u8d1f\u6837\u672c\u6781\u5ea6\u4e0d\u5747\u8861）\u3002<\/p>\n
\u2461 \u5b9a\u4f4d\u635f\u5931（localization）：L1 \u635f\u5931 + GIoU \u635f\u5931\u7684\u7ec4\u5408\u6846\u56de\u5f52，\u7ea6\u675f\u9884\u6d4b\u6846\u4e0e\u771f\u503c\u6846\u7684\u51e0\u4f55\u8d34\u5408\u5ea6\u3002<\/p>\n
\u603b\u635f\u5931 = \u5bf9\u9f50\u635f\u5931 + \u03bb \u00d7 \u5b9a\u4f4d\u635f\u5931，\u7aef\u5230\u7aef\u8054\u5408\u4f18\u5316\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u7684\u7cbe\u5999\u4e4b\u5904\u5728\u4e8e：\u5206\u7c7b\u635f\u5931\u4e0d\u518d\u662f"\u4ece\u56fa\u5b9a\u7c7b\u522b\u8868\u91cc\u6311\u4e00\u4e2a"，\u800c\u662f"\u5728\u63d0\u793a\u6587\u672c\u91cc\u627e\u5bf9\u8bcd"\u2014\u2014\u540c\u4e00\u4e2a\u635f\u5931\u51fd\u6570\u65e2\u9002\u7528\u4e8e\u4eba\u5de5\u6807\u6ce8\u7684\u68c0\u6d4b\u6570\u636e（\u6846-\u77ed\u8bed\u81ea\u7136\u5bf9\u5e94），\u4e5f\u9002\u7528\u4e8e\u56fe\u6587\u5bf9\u6570\u636e（Caption \u4e2d\u7684\u540d\u8bcd\u77ed\u8bed\u5373"\u514d\u8d39"\u7684\u7c7b\u522b\u8bcd），\u8fd9\u6b63\u662f GLIP \u80fd\u540c\u65f6\u5403\u4e0b\u4e24\u7c7b\u5f02\u6784\u6570\u636e\u3001\u5b9e\u73b0 27M \u89c4\u6a21\u9884\u8bad\u7ec3\u7684\u6839\u672c\u539f\u56e0\u3002<\/p>\n<\/p>\n
\u4e94\u3001\u7b2c 19 \u96c6：GLIP vs YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4<\/h4>\n<\/p>\n
\n\u7ef4\u5ea6YOLOv \u7cfb\u5217DETRYOLOSGLIP<\/tr>\n \n\n\u9aa8\u5e72<\/td>\n CNN<\/td>\n CNN + Transformer<\/td>\n \u88f8 ViT<\/td>\n DyHead + BERT \u8de8\u6a21\u6001<\/td>\n<\/tr>\n \n\u7c7b\u522b\u7cfb\u7edf<\/td>\n \u56fa\u5b9a<\/td>\n \u56fa\u5b9a<\/td>\n \u56fa\u5b9a<\/td>\n \u5f00\u653e\u8bcd\u6c47<\/td>\n<\/tr>\n \n\u8f93\u5165<\/td>\n \u4ec5\u56fe\u50cf<\/td>\n \u4ec5\u56fe\u50cf<\/td>\n \u4ec5\u56fe\u50cf<\/td>\n \u56fe\u50cf + \u6587\u672c\u63d0\u793a<\/td>\n<\/tr>\n \n\u540e\u5904\u7406<\/td>\n NMS<\/td>\n \u65e0（\u5308\u7259\u5229\u5339\u914d）<\/td>\n \u65e0<\/td>\n \u65e0<\/td>\n<\/tr>\n \n\u65b0\u7c7b\u522b\u8fc1\u79fb<\/td>\n \u9700\u91cd\u8bad<\/td>\n \u9700\u91cd\u8bad<\/td>\n \u9700\u91cd\u8bad<\/td>\n \u6539\u63d0\u793a\u8bcd\u5373\u53ef，\u96f6\u6837\u672c<\/td>\n<\/tr>\n \n\u4ee3\u8868\u610f\u4e49<\/td>\n CNN \u5b9e\u65f6\u68c0\u6d4b\u5dc5\u5cf0<\/td>\n \u7aef\u5230\u7aef\u68c0\u6d4b\u5f00\u5c71<\/td>\n \u67b6\u6784\u901a\u7528\u6027\u8bc1\u660e<\/td>\n \u68c0\u6d4b\u8303\u5f0f\u9769\u65b0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\u7ed3\u8bba：\u4ece YOLOv \u5230 DETR \/ YOLOS \u662f\u67b6\u6784\u5c42\u7684\u6f14\u8fdb（CNN \u2192 Transformer\u3001\u624b\u5de5\u7ec4\u4ef6 \u2192 \u7aef\u5230\u7aef），\u4ece\u5b83\u4eec\u5230 GLIP \u662f\u8303\u5f0f\u5c42\u7684\u8dc3\u8fc1（\u5c01\u95ed\u7c7b\u522b \u2192 \u5f00\u653e\u8bcd\u6c47\u3001\u89c6\u89c9\u72ec\u594f \u2192 \u8bed\u8a00\u5f15\u5bfc）\u3002\u56db\u4e2a\u6a21\u578b\u6784\u6210\u4e00\u6761\u5b8c\u6574\u7684\u6280\u672f\u8c31\u7cfb，\u4e5f\u662f\u7406\u89e3\u540e\u7eed\u5f00\u653e\u8bcd\u6c47\u5de5\u4f5c\u7684\u94a5\u5319\u3002<\/p>\n
\u516d\u3001\u8bfe\u7a0b\u8109\u7edc\u8854\u63a5<\/h4>\n
\u81f3\u6b64\u4e94\u4e2a\u6a21\u5757\u7684\u4e3b\u7ebf\u6e05\u6670\u53ef\u5faa：ViT（\u89c6\u89c9 Token \u5316）\u2192 CLIP（\u56fe\u50cf\u7ea7\u56fe\u6587\u5bf9\u9f50）\u2192 BLIP\/BLIP-2\/LLaVA（\u89c6\u89c9\u8fdb LLM \u7edf\u4e00\u63a8\u7406）\u2192 SAM（\u7c7b\u522b\u65e0\u5173\u5206\u5272）\u2192 GLIP（\u5bf9\u8c61\u7ea7\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b）\u3002GLIP \u4e0e SAM \u7684\u7ec4\u5408\u5373 Grounded-SAM（GLIP \u7528\u8bed\u8a00\u6307\u5b9a\u76ee\u6807\u5e76\u51fa\u6846 \u2192 SAM \u6cbf\u6846\u7cbe\u7ec6\u5206\u5272），\u662f\u5f53\u524d\u6700\u6d41\u884c\u7684\u5f00\u6e90\u5f00\u653e\u8bcd\u6c47\u5206\u5272\u6d41\u6c34\u7ebf\u4e4b\u4e00，\u4e5f\u4e3a\u540e\u7eed Stable Diffusion \u751f\u6210\u6a21\u5757（\u5c40\u90e8\u91cd\u7ed8\u9700\u5148\u5b9a\u4f4d\/\u5206\u5272）\u57cb\u4e0b\u4f0f\u7b14\u3002<\/p>\n
\u4e03\u3001\u5b66\u4e60\u8981\u70b9\u56de\u987e<\/h4>\n\n- \u68c0\u6d4b Transformer \u5316\u4e24\u652f\u67f1：DETR \u7684 object query + \u5308\u7259\u5229\u5339\u914d；YOLOS \u7684\u88f8 ViT \u901a\u7528\u6027\u8bc1\u660e；<\/li>\n
- GLIP \u4e00\u53e5\u8bdd：\u68c0\u6d4b = "\u56fe\u50cf\u533a\u57df \u00d7 \u63d0\u793a\u77ed\u8bed"\u7684\u5bf9\u9f50\u95ee\u9898，\u8bcd\u6c47\u8868\u7531\u6587\u672c\u5f00\u653e；<\/li>\n
- \u5173\u952e\u67b6\u6784：DyHead + BERT + \u591a\u5c42\u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408（\u533a\u522b\u4e8e CLIP \u7684 late fusion）；<\/li>\n
- \u635f\u5931 = \u533a\u57df-\u8bcd token \u7ea7\u5bf9\u9f50\u5bf9\u6bd4\u635f\u5931 + (L1 + GIoU) \u5b9a\u4f4d\u635f\u5931，\u5308\u7259\u5229\u5339\u914d\u914d\u5bf9\u540e\u8054\u5408\u4f18\u5316；<\/li>\n
- \u6570\u636e\u914d\u65b9：3M \u4eba\u5de5\u6807\u6ce8 + 24M \u56fe\u6587\u5bf9\u81ea\u8bad\u7ec3\u4f2a\u6846；<\/li>\n
- \u52a8\u624b\u5efa\u8bae：Hugging Face \u76f4\u63a5\u52a0\u8f7d GLIP \u7cfb checkpoint，\u6539\u4e00\u884c\u63d0\u793a\u8bcd\u5373\u53ef\u4f53\u9a8c\u96f6\u6837\u672c\u68c0\u6d4b，\u518d\u8fdb\u9636\u7ec4\u5408 Grounded-SAM \u6d41\u6c34\u7ebf\u3002<\/li>\n<\/ul>\n
\n\u516b\u3001GLIP \u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218<\/h4>\n
\u524d\u9762\u51e0\u8282\u4ece\u539f\u7406\u4e0a\u7406\u89e3\u4e86 GLIP \u5982\u4f55\u628a\u76ee\u6807\u68c0\u6d4b\u91cd\u6784\u4e3a\u77ed\u8bed\u5b9a\u4f4d，\u672c\u8282\u7ed9\u51fa\u53ef\u76f4\u63a5\u8fd0\u884c\u7684\u5b9e\u6218\u4ee3\u7801，\u5e2e\u52a9\u8bfb\u8005\u5728 Hugging Face \u4e0a\u5feb\u901f\u4f53\u9a8c\u96f6\u6837\u672c\u68c0\u6d4b，\u5e76\u8fdb\u9636\u7ec4\u5408 Grounded-SAM \u5b8c\u6210\u4ece\u51fa\u6846\u5230\u5206\u5272\u7684\u5b8c\u6574\u843d\u5730\u3002<\/p>\n
1. \u4f7f\u7528 Hugging Face transformers \u52a0\u8f7d GLIP \u6a21\u578b<\/p>\n
transformers \u5e93\u63d0\u4f9b\u4e86 AutoModelForZeroShotObjectDetection \u63a5\u53e3，\u53ef\u7edf\u4e00\u52a0\u8f7d GLIP \u4e0e Grounding DINO \u7b49\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u6a21\u578b\u3002\u4e0b\u9762\u4ee5 GLIP \u4e3a\u4f8b\u7ed9\u51fa\u5b8c\u6574\u63a8\u7406\u6d41\u7a0b：<\/p>\n
from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
\nfrom PIL import Image
\nimport torch
\nimport matplotlib.pyplot as plt
\nimport matplotlib.patches as patches
\n1. \u52a0\u8f7d\u6a21\u578b\u4e0e\u5904\u7406\u5668（\u9996\u6b21\u8fd0\u884c\u4f1a\u81ea\u52a8\u4e0b\u8f7d checkpoint）
\nmodel_id = "microsoft\/grounding-dino-base" # \u4e5f\u53ef\u6362\u6210 GLIP \u7cfb checkpoint
\nprocessor = AutoProcessor.from_pretrained(model_id)
\nmodel = AutoModelForZeroShotObjectDetection.from_pretrained(model_id)
\nmodel.eval()
\n2. \u8bfb\u53d6\u56fe\u50cf\u5e76\u6784\u9020\u6587\u672c\u63d0\u793a（\u5404\u7c7b\u522b\u7528\u53e5\u53f7\u5206\u9694）
\nimage = Image.open("demo.jpg").convert("RGB")
\ntext = "cat . dog . person ."
\n3. \u8f93\u5165\u9884\u5904\u7406：\u56fe\u50cf\u7f29\u653e + \u6587\u672c token \u5316，\u7edf\u4e00\u9001\u5165\u6a21\u578b
\ninputs = processor(images=image, text=text, return_tensors="pt")
\n4. \u524d\u5411\u63a8\u7406，\u5f97\u5230\u9884\u6d4b\u6846\u3001\u7c7b\u522b\u4e0e\u7f6e\u4fe1\u5ea6
\nwith torch.no_grad():
\noutputs = model(**inputs)
\n5. \u9608\u503c\u8fc7\u6ee4 + \u6846\u5750\u6807\u8f6c\u6362（\u4ece\u6a21\u578b\u5750\u6807\u8fd8\u539f\u5230\u539f\u56fe\u50cf\u7d20\u5750\u6807）
\nresults = processor.post_process_grounded_object_detection(
\noutputs,
\ninputs.input_ids,
\nbox_threshold=0.35, # \u6846\u7f6e\u4fe1\u5ea6\u9608\u503c，\u4f4e\u4e8e\u6b64\u503c\u4e22\u5f03
\ntext_threshold=0.25, # \u6587\u672c\u5bf9\u9f50\u9608\u503c，\u63a7\u5236\u7c7b\u522b\u5339\u914d\u4e25\u683c\u5ea6
\ntarget_sizes=[image.size[::-1]] # \u8fd8\u539f\u5230\u539f\u56fe\u5c3a\u5bf8
\n)[0]
\n6. \u53ef\u89c6\u5316：\u5728\u539f\u56fe\u4e0a\u7ed8\u5236\u68c0\u6d4b\u6846\u4e0e\u6807\u7b7e
\nfig, ax = plt.subplots(1, 1, figsize=(12, 8))
\nax.imshow(image)
\nfor box, score, label in zip(results["boxes"], results["scores"], results["labels"]):
\nx1, y1, x2, y2 = box.tolist()
\nrect = patches.Rectangle((x1, y1), x2 – x1, y2 – y1,
\nlinewidth=2, edgecolor="red", facecolor="none")
\nax.add_patch(rect)
\nax.text(x1, y1 – 5, f"{label} {score:.2f}",
\ncolor="red", fontsize=12, bbox=dict(facecolor="white", alpha=0.8))
\nplt.axis("off")
\nplt.show() <\/p>\n
\u5173\u952e\u6b65\u9aa4\u8bf4\u660e：\u7b2c 3 \u6b65\u7684 processor \u8d1f\u8d23\u628a\u56fe\u50cf\u7f29\u653e\u5230\u6a21\u578b\u8f93\u5165\u5c3a\u5bf8\u3001\u628a\u6587\u672c\u63d0\u793a\u7f16\u7801\u4e3a token \u5e8f\u5217；\u7b2c 5 \u6b65\u7684 box_threshold \u4e0e text_threshold \u5206\u522b\u63a7\u5236\u6846\u7f6e\u4fe1\u5ea6\u4e0e\u6587\u672c\u5bf9\u9f50\u7684\u8fc7\u6ee4\u5f3a\u5ea6，\u503c\u8d8a\u4f4e\u53ec\u56de\u8d8a\u591a\u3001\u8bef\u68c0\u4e5f\u8d8a\u591a；target_sizes \u7528\u4e8e\u628a\u6a21\u578b\u8f93\u51fa\u7684\u5f52\u4e00\u5316\u5750\u6807\u8fd8\u539f\u4e3a\u539f\u56fe\u50cf\u7d20\u5750\u6807，\u4fbf\u4e8e\u76f4\u63a5\u7ed8\u5236\u3002<\/p>\n
2. \u7ec4\u5408 Grounded-SAM：GLIP \u51fa\u6846 + SAM \u5206\u5272<\/p>\n
GLIP \u8d1f\u8d23\u7528\u8bed\u8a00\u6307\u5b9a\u76ee\u6807\u5e76\u8f93\u51fa\u8fb9\u754c\u6846，SAM \u518d\u6cbf\u6846\u5185\u533a\u57df\u505a\u7cbe\u7ec6\u5206\u5272，\u4e8c\u8005\u7ec4\u5408\u5373 Grounded-SAM \u6d41\u6c34\u7ebf\u3002\u7b80\u5316\u6d41\u7a0b\u5982\u4e0b：<\/p>\n
from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
\nfrom segment_anything import sam_model_registry, SamPredictor
\nfrom PIL import Image
\nimport numpy as np
\n\u7b2c\u4e00\u6b65：GLIP \u51fa\u6846（\u590d\u7528\u4e0a\u4e00\u8282\u7684\u52a0\u8f7d\u903b\u8f91）
\ndet_processor = AutoProcessor.from_pretrained("microsoft\/grounding-dino-base")
\ndet_model = AutoModelForZeroShotObjectDetection.from_pretrained("microsoft\/grounding-dino-base")
\nimage = Image.open("demo.jpg").convert("RGB")
\ntext = "cat . dog . person ."
\ninputs = det_processor(images=image, text=text, return_tensors="pt")
\noutputs = det_model(**inputs)
\nboxes = det_processor.post_process_grounded_object_detection(
\noutputs, inputs.input_ids,
\nbox_threshold=0.35, text_threshold=0.25,
\ntarget_sizes=[image.size[::-1]]
\n)[0]["boxes"].tolist()
\n\u7b2c\u4e8c\u6b65：SAM \u6cbf\u6846\u7cbe\u7ec6\u5206\u5272
\nsam = sam_model_registry"vit_h".to("cuda")
\npredictor = SamPredictor(sam)
\npredictor.set_image(np.array(image))
\nfor box in boxes:
\nx1, y1, x2, y2 = [int(v) for v in box]
\n# \u628a\u68c0\u6d4b\u6846\u4f5c\u4e3a SAM \u7684\u63d0\u793a\u6846，\u751f\u6210\u6846\u5185\u524d\u666f\u63a9\u7801
\nmasks, scores, _ = predictor.predict(
\nbox=np.array([x1, y1, x2, y2]),
\nmultimask_output=True
\n)
\nbest_mask = masks[scores.argmax()] # \u53d6\u7f6e\u4fe1\u5ea6\u6700\u9ad8\u7684\u63a9\u7801
\n# \u6b64\u5904\u53ef\u5c06 best_mask \u4fdd\u5b58\u4e3a PNG \u6216\u53e0\u52a0\u5230\u539f\u56fe\u7528\u4e8e\u540e\u7eed\u7f16\u8f91 <\/p>\n
\u8fd9\u6bb5\u6d41\u7a0b\u628a GLIP \u7684\u5f00\u653e\u8bcd\u6c47\u5b9a\u4f4d\u80fd\u529b\u4e0e SAM \u7684\u7c7b\u522b\u65e0\u5173\u5206\u5272\u80fd\u529b\u4e32\u8054\u8d77\u6765：GLIP \u5148\u56de\u7b54"\u56fe\u4e2d\u54ea\u91cc\u6709 cat \/ dog \/ person"，SAM \u518d\u56de\u7b54"\u8fd9\u4e9b\u76ee\u6807\u5404\u81ea\u7684\u7cbe\u786e\u8f6e\u5ed3\u662f\u4ec0\u4e48"\u3002\u5728\u5f00\u653e\u8bcd\u6c47\u56fe\u50cf\u7f16\u8f91\u573a\u666f\u4e2d，\u8fd9\u4e00\u7ec4\u5408\u975e\u5e38\u5b9e\u7528\u2014\u2014\u4f8b\u5982\u7528\u6237\u8f93\u5165"\u628a\u56fe\u4e2d\u7684\u67f4\u72ac\u6362\u6210\u67ef\u57fa"，\u5373\u53ef\u5148\u7528 GLIP \u5b9a\u4f4d"\u67f4\u72ac"\u533a\u57df，\u518d\u7528 SAM \u5206\u5272\u51fa\u8be5\u533a\u57df\u63a9\u7801，\u6700\u540e\u4ea4\u7ed9 Stable Diffusion \u505a\u5c40\u90e8\u91cd\u7ed8，\u5b9e\u73b0\u76ee\u6807\u7ea7\u3001\u8bed\u4e49\u53ef\u63a7\u7684\u56fe\u50cf\u7f16\u8f91\u3002<\/p>\n
\u516b\u3001\u5b66\u4e60\u8def\u7ebf\u56fe<\/h4>\n
\u9488\u5bf9\u4e0d\u540c\u57fa\u7840\u7684\u5b66\u4e60\u8005，\u5efa\u8bae\u6309\u4ee5\u4e0b\u8def\u5f84\u5faa\u5e8f\u6e10\u8fdb\u5730\u638c\u63e1 GLIP \u53ca\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u6280\u672f：<\/p>\n
\n- \u5165\u95e8\u8def\u7ebf（0\u20132 \u5468）：\u5148\u638c\u63e1 Transformer \u57fa\u7840\u4e0e ViT \u539f\u7406，\u518d\u5b66\u4e60 DETR \u7684\u7aef\u5230\u7aef\u68c0\u6d4b\u8303\u5f0f，\u7406\u89e3 object query \u4e0e\u5308\u7259\u5229\u5339\u914d\u673a\u5236；<\/li>\n
- \u8fdb\u9636\u8def\u7ebf（2\u20134 \u5468）：\u6df1\u5165 YOLOS \u7684\u88f8 ViT \u68c0\u6d4b\u5b9e\u9a8c，\u7406\u89e3\u68c0\u6d4b\u80fd\u529b\u7684\u6d8c\u73b0\u673a\u5236；\u968f\u540e\u7cfb\u7edf\u5b66\u4e60 GLIP \u7684\u67b6\u6784\u4e09\u4ef6\u5957（DyHead + BERT + \u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408），\u638c\u63e1\u77ed\u8bed\u5b9a\u4f4d\u7684\u6838\u5fc3\u601d\u60f3；<\/li>\n
- \u5b9e\u6218\u8def\u7ebf（4\u20136 \u5468）：\u5728 Hugging Face \u4e0a\u52a0\u8f7d GLIP \u7cfb checkpoint，\u52a8\u624b\u5b9e\u8df5\u96f6\u6837\u672c\u68c0\u6d4b；\u8fdb\u9636\u7ec4\u5408 Grounded-SAM \u6d41\u6c34\u7ebf，\u5b8c\u6210\u4ece\u8bed\u8a00\u6307\u5b9a\u76ee\u6807\u5230\u7cbe\u7ec6\u5206\u5272\u7684\u5b8c\u6574\u843d\u5730；<\/li>\n
- \u7814\u7a76\u8def\u7ebf（6 \u5468\u4ee5\u4e0a）：\u7cbe\u8bfb GLIP \u8bba\u6587\u4e0e\u6e90\u7801，\u590d\u73b0\u635f\u5931\u51fd\u6570\u4e0e\u81ea\u8bad\u7ec3\u6570\u636e\u914d\u65b9，\u5c1d\u8bd5\u5728\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u4e0a\u5fae\u8c03，\u63a2\u7d22\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u7684\u8fb9\u754c\u4e0e\u6539\u8fdb\u65b9\u5411\u3002<\/li>\n<\/ul>\n
SEO \u5173\u952e\u8bcd：GLIP\u3001\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u3001\u591a\u6a21\u6001\u5927\u6a21\u578b\u3001phrase grounding\u3001YOLOS\u3001DETR \u539f\u7406\u3001DyHead\u3001\u96f6\u6837\u672c\u68c0\u6d4b\u3001GLIP \u635f\u5931\u51fd\u6570\u3001Grounded-SAM \u6559\u7a0b\u3001\u76ee\u6807\u68c0\u6d4b Transformer\u3001\u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408\u3002<\/p>\n
\u4e5d\u3001\u603b\u7ed3\u4e0e\u53c2\u8003\u8d44\u6599<\/h4>\n
\u672c\u6587\u7cfb\u7edf\u68b3\u7406\u4e86\u5362\u83c1\u535a\u58eb\u300a\u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b\u300b\u7b2c\u4e94\u6a21\u5757\u300aGLIP \u6a21\u578b\u8be6\u89e3\u300b（\u7b2c 16\u201319 \u96c6）\u7684\u6838\u5fc3\u5185\u5bb9：\u4ece DETR\u3001YOLOS \u7684\u68c0\u6d4b Transformer \u5316\u8bb2\u8d77，\u6df1\u5165\u5256\u6790 GLIP \u5982\u4f55\u628a\u76ee\u6807\u68c0\u6d4b\u91cd\u6784\u4e3a\u77ed\u8bed\u5b9a\u4f4d\u4efb\u52a1，\u5b9e\u73b0\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u3002GLIP \u901a\u8fc7 DyHead + BERT + \u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408\u7684\u67b6\u6784\u8bbe\u8ba1，\u914d\u5408 27M \u9884\u8bad\u7ec3\u6570\u636e\u914d\u65b9\u4e0e\u8054\u5408\u635f\u5931\u51fd\u6570，\u5728\u96f6\u6837\u672c\u68c0\u6d4b\u4e0a\u53d6\u5f97\u4e86\u8d85\u8d8a\u6709\u76d1\u7763\u57fa\u7ebf\u7684\u6210\u7ee9，\u6807\u5fd7\u7740\u68c0\u6d4b\u8303\u5f0f\u4ece\u5c01\u95ed\u7c7b\u522b\u5230\u5f00\u653e\u8bcd\u6c47\u7684\u5173\u952e\u8dc3\u8fc1\u3002<\/p>\n
\u4ee5\u4e0b\u662f\u5b98\u65b9\u6587\u6863\u4e0e\u63a8\u8350\u9605\u8bfb\u8d44\u6599：<\/p>\n
\n- GLIP \u8bba\u6587：Grounded Language-Image Pre-training（CVPR 2022，\u5fae\u8f6f\u7814\u7a76\u9662）<\/li>\n
- Microsoft Research \u9879\u76ee\u9875：Object Detection in the Wild via Grounded Language-Image Pre-training<\/li>\n
- Hugging Face \u6a21\u578b\u5e93：\u641c\u7d22 GLIP \u5373\u53ef\u627e\u5230\u5b98\u65b9 checkpoint \u4e0e\u4f7f\u7528\u793a\u4f8b，\u652f\u6301\u96f6\u6837\u672c\u68c0\u6d4b\u5feb\u901f\u4f53\u9a8c<\/li>\n
- Grounded-SAM \u9879\u76ee：IDEA-Research\/Grounded-Segment-Anything，GLIP \u4e0e SAM \u7ec4\u5408\u7684\u5f00\u653e\u8bcd\u6c47\u5206\u5272\u6d41\u6c34\u7ebf<\/li>\n
- \u8bfe\u7a0b\u9009\u96c6：\u5362\u83c1\u535a\u58eb\u300a\u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b\u300b，\u7b2c 16\u201319 \u96c6\u5bf9\u5e94\u672c\u6a21\u5757\u5185\u5bb9<\/li>\n<\/ul>\n
<\/h3>\n
\u4e3b\u8981\u6765\u6e90：GLIP \u8bba\u6587 \u94fe\u63a5\u3001Microsoft Research \u9879\u76ee\u9875 \u94fe\u63a5\u3001GLIP \u6280\u672f\u89e3\u8bfb \u94fe\u63a5\u3001\u8bfe\u7a0b\u9009\u96c6 \u94fe\u63a5<\/p><\/p>\n","protected":false},"excerpt":{"rendered":"
\u76ee\u5f55：
\n\u4e00\u3001\u6a21\u5757\u5b9a\u4f4d：\u4ece\\”\u5206\u5272\u4e00\u5207\\”\u5230\\”\u68c0\u6d4b\u4e00\u5207\\”\u4e8c\u3001\u7b2c 16 \u96c6：YOLOS \u4e0e DETR\u2014\u2014\u68c0\u6d4b\u7684 Transformer \u5316\u4e09\u3001\u7b2c 17 \u96c6：GLIP \u539f\u7406\u2014\u2014\u68c0\u6d4b\u5373\u77ed\u8bed\u5b9a\u4f4d\u56db\u3001\u7b2c 18 \u96c6：GLIP \u635f\u5931\u51fd\u6570\u4e94\u3001\u7b2c 19 \u96c6：GLIP vs YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4\u516d\u3001\u8bfe\u7a0b\u8109\u7edc\u8854\u63a5\u4e03\u3001\u5b66\u4e60\u8981\u70b9\u56de\u987e\u516b\u3001\u5b66\u4e60\u8def\u7ebf\u56fe\u4e5d\u3001\u603b\u7ed3\u4e0e\u53c2\u8003\u8d44\u6599
\n\u6458\u8981：\u672c\u6587\u7cfb\u7edf\u68b3<\/p>\n","protected":false},"author":2,"featured_media":111726,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[11653,50,1665,523],"topic":[],"class_list":["post-111727","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-11653","tag-50","tag-1665","tag-523"],"yoast_head":"\n
GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/111727.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u76ee\u5f55： \u4e00\u3001\u6a21\u5757\u5b9a\u4f4d：\u4ece"\u5206\u5272\u4e00\u5207"\u5230"\u68c0\u6d4b\u4e00\u5207"\u4e8c\u3001\u7b2c 16 \u96c6：YOLOS \u4e0e DETR\u2014\u2014\u68c0\u6d4b\u7684 Transformer \u5316\u4e09\u3001\u7b2c 17 \u96c6：GLIP \u539f\u7406\u2014\u2014\u68c0\u6d4b\u5373\u77ed\u8bed\u5b9a\u4f4d\u56db\u3001\u7b2c 18 \u96c6：GLIP \u635f\u5931\u51fd\u6570\u4e94\u3001\u7b2c 19 \u96c6：GLIP vs YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4\u516d\u3001\u8bfe\u7a0b\u8109\u7edc\u8854\u63a5\u4e03\u3001\u5b66\u4e60\u8981\u70b9\u56de\u987e\u516b\u3001\u5b66\u4e60\u8def\u7ebf\u56fe\u4e5d\u3001\u603b\u7ed3\u4e0e\u53c2\u8003\u8d44\u6599 \u6458\u8981：\u672c\u6587\u7cfb\u7edf\u68b3\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/111727.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-01T08:50:23+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261001085022-6abe1ece5dac9.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/111727.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/111727.html\",\"name\":\"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-01T08:50:23+00:00\",\"dateModified\":\"2026-10-01T08:50:23+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/111727.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/111727.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/111727.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/111727.html","og_locale":"zh_CN","og_type":"article","og_title":"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u76ee\u5f55： \u4e00\u3001\u6a21\u5757\u5b9a\u4f4d：\u4ece\"\u5206\u5272\u4e00\u5207\"\u5230\"\u68c0\u6d4b\u4e00\u5207\"\u4e8c\u3001\u7b2c 16 \u96c6：YOLOS \u4e0e DETR\u2014\u2014\u68c0\u6d4b\u7684 Transformer \u5316\u4e09\u3001\u7b2c 17 \u96c6：GLIP \u539f\u7406\u2014\u2014\u68c0\u6d4b\u5373\u77ed\u8bed\u5b9a\u4f4d\u56db\u3001\u7b2c 18 \u96c6：GLIP \u635f\u5931\u51fd\u6570\u4e94\u3001\u7b2c 19 \u96c6：GLIP vs YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4\u516d\u3001\u8bfe\u7a0b\u8109\u7edc\u8854\u63a5\u4e03\u3001\u5b66\u4e60\u8981\u70b9\u56de\u987e\u516b\u3001\u5b66\u4e60\u8def\u7ebf\u56fe\u4e5d\u3001\u603b\u7ed3\u4e0e\u53c2\u8003\u8d44\u6599 \u6458\u8981：\u672c\u6587\u7cfb\u7edf\u68b3","og_url":"https:\/\/www.wsisp.com\/helps\/111727.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-01T08:50:23+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261001085022-6abe1ece5dac9.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/111727.html","url":"https:\/\/www.wsisp.com\/helps\/111727.html","name":"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-01T08:50:23+00:00","dateModified":"2026-10-01T08:50:23+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/111727.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/111727.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/111727.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/111727","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=111727"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/111727\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/111726"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=111727"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=111727"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=111727"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=111727"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
GLIP \u5f00\u653e\u8bcd\u6c47\u76ee\u6807\u68c0\u6d4b\u7cbe\u8bb2（\u542b YOLOS\/DETR \u539f\u7406） | \u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b \u7b2c 16\u201319 \u96c6<\/h3>\n- \n
- \u6587\u4ef6：artifacts\/glip-model-guide.html（\u6697\u8272\u8bbe\u8ba1，\u542b\u5206\u96c6\u5bfc\u822a\u3001GLIP \u67b6\u6784\u6d41\u7a0b\u56fe\u3001\u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408\u793a\u610f\u3001\u635f\u5931\u51fd\u6570\u516c\u5f0f\u5757\u3001\u56db\u6a21\u578b\u5bf9\u6bd4\u8868\u3001\u8bfe\u7a0b\u8109\u7edc\u65f6\u95f4\u7ebf，\u4ee3\u7801\u7ea7\u8d28\u68c0 PASS）<\/li>\n<\/ul>\n
\u5206\u96c6\u7ed3\u6784\u5b9e\u8bc1（\u6765\u6e90：\u8bfe\u7a0b\u9009\u96c6 \u94fe\u63a5）：<\/p>\n
\n
\u96c6\u6570\u6807\u9898\u65f6\u957f<\/tr>\n \n\n \u7b2c 16 \u96c6<\/td>\n GLIP \u6a21\u578b\u8be6\u89e3(1) YOLOS \u548c DETR \u6a21\u578b\u539f\u7406\u89e3\u6790<\/td>\n 11.85 \u5206\u949f<\/td>\n<\/tr>\n \n \u7b2c 17 \u96c6<\/td>\n GLIP \u6a21\u578b\u8be6\u89e3(2) GLIP \u6a21\u578b\u539f\u7406\u89e3\u6790<\/td>\n 11.53 \u5206\u949f<\/td>\n<\/tr>\n \n \u7b2c 18 \u96c6<\/td>\n GLIP \u6a21\u578b\u8be6\u89e3(3) GLIP \u635f\u5931\u51fd\u6570<\/td>\n 12.43 \u5206\u949f<\/td>\n<\/tr>\n \n \u7b2c 19 \u96c6<\/td>\n GLIP \u6a21\u578b\u8be6\u89e3(4) GLIP \u4e0e YOLOS\/DETR\/YOLOv \u5bf9\u6bd4<\/td>\n 9.03 \u5206\u949f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n\u8be6\u7ec6\u603b\u7ed3（\u7ea6 5200 \u5b57）<\/h3>\n
\u4e00\u3001\u6a21\u5757\u5b9a\u4f4d：\u4ece"\u5206\u5272\u4e00\u5207"\u5230"\u68c0\u6d4b\u4e00\u5207"<\/h4>\n
\u672c\u6a21\u5757\u4e3a\u5362\u83c1\u535a\u58eb\u300a\u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b\u300b\u7b2c\u4e94\u6a21\u5757，\u5171 4 \u96c6，\u603b\u65f6\u957f\u7ea6 44.8 \u5206\u949f\u3002\u4e0a\u4e00\u6a21\u5757 SAM \u89e3\u51b3\u4e86"\u7c7b\u522b\u65e0\u5173\u7684\u5206\u5272"，\u4f46\u7559\u4e0b\u4e00\u4e2a\u660e\u786e\u7f3a\u53e3：SAM \u53ea\u753b\u8f6e\u5ed3\u3001\u4e0d\u8ba4\u7c7b\u522b\u3002\u672c\u6a21\u5757\u8bb2\u89e3\u7684 GLIP（Grounded Language-Image Pre-training，\u5fae\u8f6f\u7814\u7a76\u9662，CVPR 2022）\u6070\u597d\u8865\u4e0a\u8fd9\u5757\u62fc\u56fe\u2014\u2014\u7528\u81ea\u7136\u8bed\u8a00\u5b9a\u4e49"\u8981\u68c0\u6d4b\u4ec0\u4e48"，\u628a\u5f00\u653e\u8bcd\u6c47\u80fd\u529b\u5e26\u5230\u5bf9\u8c61\u7ea7\u5b9a\u4f4d\u4efb\u52a1\u3002<\/p>\n
\u8bfe\u7a0b\u7f16\u6392\u9887\u5177\u5320\u5fc3：\u7b2c 16 \u96c6\u5e76\u4e0d\u76f4\u63a5\u8bb2 GLIP，\u800c\u662f\u5148\u94fa\u57ab\u4e24\u4ee3\u68c0\u6d4b Transformer（DETR\u3001YOLOS），\u56e0\u4e3a GLIP \u7684\u68c0\u6d4b\u9aa8\u5e72\u6b63\u662f\u5efa\u7acb\u5728\u8fd9\u6761\u6280\u672f\u7ebf\u4e0a\u3002\u7406\u89e3\u4e86"\u68c0\u6d4b\u5982\u4f55 Transformer \u5316"，\u624d\u80fd\u7406\u89e3 GLIP \u5728\u5176\u4e0a\u505a\u7684\u8303\u5f0f\u9769\u65b0\u3002<\/p>\n
\u4e8c\u3001\u7b2c 16 \u96c6：YOLOS \u4e0e DETR\u2014\u2014\u68c0\u6d4b\u7684 Transformer \u5316<\/h4>\n
DETR（2020，Facebook AI）：\u9996\u4e2a\u7aef\u5230\u7aef\u76ee\u6807\u68c0\u6d4b Transformer\u3002\u6d41\u7a0b\u4e3a：CNN \u9aa8\u5e72\u63d0\u53d6\u56fe\u50cf\u7279\u5f81 \u2192 Transformer \u7f16\u7801\u5668\u505a\u5168\u5c40\u5efa\u6a21 \u2192 \u89e3\u7801\u5668\u7528\u4e00\u7ec4\u56fa\u5b9a\u6570\u91cf（\u5982 100 \u4e2a）\u7684 object query \u5e76\u884c"\u8be2\u95ee"\u56fe\u50cf\u4e2d\u6bcf\u4e2a\u76ee\u6807\u7684\u4f4d\u7f6e \u2192 \u8f93\u51fa\u8fb9\u754c\u6846\u4e0e\u7c7b\u522b\u3002\u5b83\u6709\u4e24\u5927\u9769\u547d\u6027\u8bbe\u8ba1：<\/p>\n
- \u629b\u5f03\u951a\u6846（anchor）\u3002\u4f20\u7edf\u68c0\u6d4b\u5668（Faster R-CNN\u3001YOLO）\u4f9d\u8d56\u4eba\u5de5\u8bbe\u8ba1\u7684\u5148\u9a8c\u951a\u6846，\u5c3a\u5bf8\u3001\u6bd4\u4f8b\u3001\u6570\u91cf\u90fd\u662f\u8d85\u53c2\u6570；DETR \u7684 object query \u662f\u4e00\u7ec4\u53ef\u5b66\u4e60\u7684\u4f4d\u7f6e\u67e5\u8be2\u5411\u91cf，\u6a21\u578b\u81ea\u5df1\u5b66\u4f1a"\u53bb\u54ea\u91cc\u770b"，\u4e0d\u518d\u4f9d\u8d56\u4eba\u5de5\u5148\u9a8c\u3002<\/li>\n
- \u629b\u5f03 NMS（\u975e\u6781\u5927\u503c\u6291\u5236）\u3002\u4f20\u7edf\u68c0\u6d4b\u5668\u4f1a\u5bf9\u540c\u4e00\u76ee\u6807\u8f93\u51fa\u591a\u4e2a\u91cd\u53e0\u6846，\u518d\u7528 NMS \u8d2a\u5fc3\u53bb\u91cd；DETR \u7528**\u5308\u7259\u5229\u5339\u914d（\u4e8c\u5206\u56fe\u6700\u4f18\u5339\u914d）**\u5b9e\u73b0\u9884\u6d4b\u4e0e\u771f\u503c\u7684\u4e00\u5bf9\u4e00\u5206\u914d\u2014\u2014\u6bcf\u4e2a\u771f\u503c\u6846\u53ea\u7531\u4e00\u4e2a\u9884\u6d4b\u8d1f\u8d23，\u4ece\u673a\u5236\u6839\u4e0a\u6d88\u9664\u4e86\u91cd\u590d\u6846，\u5b9e\u73b0\u4e86\u771f\u6b63\u7684\u7aef\u5230\u7aef\u3002<\/li>\n
\u4ee3\u4ef7\u4e5f\u5f88\u660e\u663e：\u6536\u655b\u6162（\u9700\u6570\u767e\u4e2a epoch）\u3001\u5c0f\u76ee\u6807\u68c0\u6d4b\u504f\u5f31\u3002\u8fd9\u4e9b\u95ee\u9898\u50ac\u751f\u4e86\u540e\u7eed\u5927\u91cf\u6539\u8fdb\u5de5\u4f5c（Deformable DETR \u7b49），\u4f46"\u7aef\u5230\u7aef + \u96c6\u5408\u9884\u6d4b"\u7684\u8303\u5f0f\u7531\u6b64\u786e\u7acb\u3002<\/p>\n
YOLOS（2021）：\u4e00\u4e2a"\u6781\u7b80\u5b9e\u9a8c"\u2014\u2014\u62ff\u4e3a\u5206\u7c7b\u9884\u8bad\u7ec3\u7684\u88f8 ViT，\u51e0\u4e4e\u4e0d\u6539\u67b6\u6784，\u53ea\u8ffd\u52a0\u5c11\u91cf\u53ef\u5b66\u4e60\u7684 [DET] token，\u5c31\u76f4\u63a5\u505a\u76ee\u6807\u68c0\u6d4b，\u4e14\u6027\u80fd\u53ef\u89c2\u3002\u5b83\u7684\u610f\u4e49\u4e0d\u5728\u5237\u699c，\u800c\u5728\u8bc1\u660e：\u68c0\u6d4b\u80fd\u529b\u53ef\u4ee5\u4ece\u7eaf\u7cb9\u7684\u5e8f\u5217\u5230\u5e8f\u5217\u5efa\u6a21\u4e2d"\u6d8c\u73b0"，\u4e0d\u9700\u8981 CNN \u7684\u5f52\u7eb3\u504f\u7f6e，\u4e5f\u4e0d\u9700\u8981\u68c0\u6d4b\u4e13\u7528\u67b6\u6784\u3002\u8fd9\u4e0e\u8bfe\u7a0b\u7b2c\u4e00\u6a21\u5757 ViT"Transformer \u53ef\u5b8c\u5168\u66ff\u4ee3 CNN \u505a\u89c6\u89c9"\u7684\u7ed3\u8bba\u4e00\u8109\u76f8\u627f，\u628a"Transformer \u901a\u7528\u6027"\u4ece\u5206\u7c7b\u63a8\u8fdb\u5230\u4e86\u5b9a\u4f4d\u4efb\u52a1\u3002<\/p>\n
\u8fd9\u4e00\u96c6\u5408\u8d77\u6765\u56de\u7b54\u7684\u95ee\u9898\u662f：\u68c0\u6d4b\u5668\u5df2\u7ecf Transformer \u5316\u4e86，\u4f46\u5b83\u4ecd\u662f"\u5c01\u95ed\u4e16\u754c"\u7684\u2014\u2014\u7c7b\u522b\u8868\u56fa\u5b9a（\u5982 COCO 80 \u7c7b），\u6362\u4efb\u52a1\u5c31\u8981\u6539\u5206\u7c7b\u5934\u3001\u91cd\u65b0\u8bad\u7ec3\u3002\u8fd9\u5c31\u662f GLIP \u8981\u89e3\u51b3\u7684\u6839\u672c\u95ee\u9898\u3002<\/p>\n
\u4e09\u3001\u7b2c 17 \u96c6：GLIP \u539f\u7406\u2014\u2014\u68c0\u6d4b\u5373\u77ed\u8bed\u5b9a\u4f4d<\/h4>\n
GLIP \u7684\u6838\u5fc3\u601d\u60f3\u7528\u4e00\u53e5\u8bdd\u6982\u62ec：\u628a\u76ee\u6807\u68c0\u6d4b\u91cd\u6784\u4e3a\u77ed\u8bed\u5b9a\u4f4d（phrase grounding）\u4efb\u52a1\u3002<\/p>\n
- \n
- \u8f93\u5165：\u56fe\u50cf + \u4e00\u6bb5\u6587\u672c\u63d0\u793a，\u5982 cat . dog . person .（\u5404\u7c7b\u522b\u7528\u53e5\u53f7\u5206\u9694\u7684\u63d0\u793a\u53e5）<\/li>\n
- \u8f93\u51fa：\u56fe\u4e2d\u6bcf\u4e2a\u4e0e\u63d0\u793a\u4e2d\u77ed\u8bed\u8bed\u4e49\u5bf9\u5e94\u7684\u76ee\u6807\u6846<\/li>\n<\/ul>\n
\u5728 GLIP \u6846\u67b6\u4e0b，\u201c\u68c0\u6d4b"\u88ab\u7edf\u4e00\u8fdb grounding \u7684\u8bed\u4e49：\u4f20\u7edf\u68c0\u6d4b\u5668\u8f93\u51fa"\u8fd9\u662f\u7c7b\u522b\u7f16\u53f7 #15\u201d，GLIP \u8f93\u51fa"\u8fd9\u4e2a\u533a\u57df\u4e0e\u63d0\u793a\u4e2d\u7684\u8bcd \u2018cat\u2019 \u5bf9\u9f50"\u3002\u5206\u7c7b\u5934\u88ab\u66ff\u6362\u4e3a\u89c6\u89c9\u533a\u57df\u7279\u5f81\u4e0e\u6587\u672c\u8bcd\u7279\u5f81\u7684\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u2014\u2014\u68c0\u6d4b\u7684"\u7c7b\u522b\u8868"\u4ece\u6b64\u53d8\u6210\u4efb\u610f\u81ea\u7136\u8bed\u8a00，\u8bcd\u6c47\u8868\u5f7b\u5e95\u5f00\u653e：\u9762\u5bf9\u8bad\u7ec3\u96c6\u4e2d\u4ece\u672a\u51fa\u73b0\u7684\u65b0\u7c7b\u522b（\u5982"\u6234\u592a\u9633\u955c\u7684\u67f4\u72ac"），\u53ea\u9700\u6539\u63d0\u793a\u8bcd，\u96f6\u6837\u672c\u5373\u53ef\u68c0\u6d4b\u3002<\/p>\n
\u67b6\u6784\u4e09\u4ef6\u5957：<\/p>\n
- \u89c6\u89c9\u7f16\u7801\u5668：DyHead\u3002\u52a8\u6001\u5934\u7ed3\u6784，\u5728\u5c3a\u5ea6\u3001\u7a7a\u95f4\u3001\u4efb\u52a1\u4e09\u4e2a\u7ef4\u5ea6\u4e0a\u505a\u6ce8\u610f\u529b，\u4ea7\u51fa\u533a\u57df\u7ea7\u89c6\u89c9\u7279\u5f81\u3002<\/li>\n
- \u6587\u672c\u7f16\u7801\u5668：BERT（\u6216 CLIP \u6587\u672c\u7f16\u7801\u5668）\u3002\u628a\u63d0\u793a\u53e5\u7f16\u7801\u4e3a\u8bcd\u7ea7\u7279\u5f81\u5e8f\u5217\u3002<\/li>\n
- \u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408（language-aware deep fusion）\u3002\u8fd9\u662f GLIP \u533a\u522b\u4e8e CLIP \u7684\u5173\u952e\u3002CLIP \u53ea\u5728\u6700\u540e\u5bf9\u5168\u5c40\u7279\u5f81\u505a\u70b9\u79ef（late fusion，\u201c\u5404\u5b66\u5404\u7684\u3001\u6700\u540e\u5bf9\u7b54\u6848\u201d），\u662f\u56fe\u50cf\u7ea7\u7684\u6d45\u5bf9\u9f50；GLIP \u5219\u5728\u4e24\u4e2a\u7f16\u7801\u5668\u7684\u591a\u4e2a\u4e2d\u95f4\u5c42\u63d2\u5165\u8de8\u6a21\u6001\u4ea4\u53c9\u6ce8\u610f\u529b\u2014\u2014\u89c6\u89c9\u5c42\u7528\u6587\u672c\u7279\u5f81\u66f4\u65b0\u81ea\u5df1（"\u7ea2\u8272"\u8fd9\u4e2a\u8bcd\u4f1a\u5b9e\u65f6\u589e\u5f3a\u7ea2\u8272\u533a\u57df\u7684\u7279\u5f81\u6743\u91cd），\u6587\u672c\u5c42\u4e5f\u7528\u89c6\u89c9\u7279\u5f81\u56de\u5199，\u53cc\u5411\u9010\u5c42\u4ea4\u6362\u4fe1\u606f，\u5b9e\u73b0\u5bf9\u8c61\u7ea7\u7684\u6df1\u5ea6\u8bed\u4e49\u8026\u5408\u3002<\/li>\n
\u9884\u8bad\u7ec3\u6570\u636e\u914d\u65b9（27M）：<\/p>\n
- \n
- 3M \u4eba\u5de5\u6807\u6ce8：Flickr30K Entities\u3001Objects365 \u7b49\u68c0\u6d4b\/\u5b9a\u4f4d\u6570\u636e，\u6846\u4e0e\u77ed\u8bed\u5929\u7136\u5bf9\u5e94，\u8d28\u91cf\u9ad8\u4f46\u89c4\u6a21\u6709\u9650；<\/li>\n
- 24M \u7f51\u7edc\u56fe\u6587\u5bf9：\u6ca1\u6709\u6846\u6807\u6ce8\u3002GLIP \u7528**\u81ea\u8bad\u7ec3（self-training）**\u65b9\u5f0f\u2014\u2014\u5148\u8bad\u51fa\u7684\u6a21\u578b\u7ed9\u56fe\u6587\u5bf9 Caption \u4e2d\u7684\u540d\u8bcd\u77ed\u8bed"\u6253\u4f2a\u6846"，\u8fc7\u6ee4\u4f4e\u7f6e\u4fe1\u5ea6\u7ed3\u679c\u540e\u56de\u7089\u8bad\u7ec3\u3002\u8fd9\u4e0e CLIP \u5229\u7528 4 \u4ebf\u56fe\u6587\u5bf9\u7684\u601d\u8def\u540c\u6e90，\u4f46\u628a\u5bf9\u9f50\u7c92\u5ea6\u4ece\u56fe\u50cf\u7ea7\u4e0b\u6c89\u5230\u4e86\u5bf9\u8c61\u7ea7\u3002<\/li>\n<\/ul>\n
\u6548\u679c：GLIP-L \u5728\u96f6\u6837\u672c COCO \u68c0\u6d4b\u4e0a\u8fbe\u5230 49.8 AP，\u8d85\u8fc7\u6709\u76d1\u7763\u8bad\u7ec3\u7684\u5f3a\u57fa\u7ebf；\u5728 13 \u4e2a\u975e\u5e38\u89c4\u76ee\u6807\u6570\u636e\u96c6\u4e0a\u5e73\u5747\u8d85\u8d8a\u6709\u76d1\u7763\u65b9\u6cd5，\u8bc1\u660e\u4e86\u5f00\u653e\u8bcd\u6c47\u8def\u7ebf\u7684\u5b9e\u7528\u4ef7\u503c\u3002<\/p>\n
\u56db\u3001\u7b2c 18 \u96c6：GLIP \u635f\u5931\u51fd\u6570<\/h4>\n
GLIP \u7684\u8bad\u7ec3\u76ee\u6807\u662f"\u5206\u7c7b\u4e0e\u5b9a\u4f4d\u7edf\u4e00"\u7684\u8054\u5408\u635f\u5931，\u7531\u4e24\u90e8\u5206\u7ec4\u6210：<\/p>\n
\u2460 \u8de8\u6a21\u6001\u5bf9\u9f50\u635f\u5931（classification as grounding）：\u5148\u6784\u5efa"\u533a\u57df \u00d7 \u8bcd"\u76f8\u4f3c\u5ea6\u77e9\u9635 S = O\u00b7P\u1d40（\u533a\u57df\u7279\u5f81 O \u4e0e\u8bcd\u7279\u5f81 P \u7684\u70b9\u79ef），\u7528\u5b83\u66ff\u4ee3\u4f20\u7edf\u7684\u56fa\u5b9a\u7c7b\u522b softmax \u5206\u7c7b\u5934\u3002\u5308\u7259\u5229\u5339\u914d\u786e\u5b9a\u9884\u6d4b\u6846\u4e0e\u771f\u503c\u77ed\u8bed\u7684\u914d\u5bf9\u540e，\u5728\u77e9\u9635\u4e0a\u505a token \u7ea7\u5bf9\u9f50：\u6b63\u6837\u672c\u5bf9（\u6846\u4e0e\u5bf9\u5e94\u77ed\u8bed\u7684\u8bcd）\u76f8\u4f3c\u5ea6\u63a8\u9ad8，\u8d1f\u6837\u672c\u5bf9\u63a8\u4f4e\u2014\u2014\u672c\u8d28\u662f\u4e00\u4e2a\u5bf9\u8c61\u7ea7\u7684\u5bf9\u6bd4\u5206\u7c7b\u635f\u5931（\u7126\u70b9\u52a0\u6743\u4ee5\u5904\u7406\u6b63\u8d1f\u6837\u672c\u6781\u5ea6\u4e0d\u5747\u8861）\u3002<\/p>\n
\u2461 \u5b9a\u4f4d\u635f\u5931（localization）：L1 \u635f\u5931 + GIoU \u635f\u5931\u7684\u7ec4\u5408\u6846\u56de\u5f52，\u7ea6\u675f\u9884\u6d4b\u6846\u4e0e\u771f\u503c\u6846\u7684\u51e0\u4f55\u8d34\u5408\u5ea6\u3002<\/p>\n
\u603b\u635f\u5931 = \u5bf9\u9f50\u635f\u5931 + \u03bb \u00d7 \u5b9a\u4f4d\u635f\u5931，\u7aef\u5230\u7aef\u8054\u5408\u4f18\u5316\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u7684\u7cbe\u5999\u4e4b\u5904\u5728\u4e8e：\u5206\u7c7b\u635f\u5931\u4e0d\u518d\u662f"\u4ece\u56fa\u5b9a\u7c7b\u522b\u8868\u91cc\u6311\u4e00\u4e2a"，\u800c\u662f"\u5728\u63d0\u793a\u6587\u672c\u91cc\u627e\u5bf9\u8bcd"\u2014\u2014\u540c\u4e00\u4e2a\u635f\u5931\u51fd\u6570\u65e2\u9002\u7528\u4e8e\u4eba\u5de5\u6807\u6ce8\u7684\u68c0\u6d4b\u6570\u636e（\u6846-\u77ed\u8bed\u81ea\u7136\u5bf9\u5e94），\u4e5f\u9002\u7528\u4e8e\u56fe\u6587\u5bf9\u6570\u636e（Caption \u4e2d\u7684\u540d\u8bcd\u77ed\u8bed\u5373"\u514d\u8d39"\u7684\u7c7b\u522b\u8bcd），\u8fd9\u6b63\u662f GLIP \u80fd\u540c\u65f6\u5403\u4e0b\u4e24\u7c7b\u5f02\u6784\u6570\u636e\u3001\u5b9e\u73b0 27M \u89c4\u6a21\u9884\u8bad\u7ec3\u7684\u6839\u672c\u539f\u56e0\u3002<\/p>\n<\/p>\n
\u4e94\u3001\u7b2c 19 \u96c6：GLIP vs YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4<\/h4>\n<\/p>\n
\n
\u7ef4\u5ea6YOLOv \u7cfb\u5217DETRYOLOSGLIP<\/tr>\n \n\n \u9aa8\u5e72<\/td>\n CNN<\/td>\n CNN + Transformer<\/td>\n \u88f8 ViT<\/td>\n DyHead + BERT \u8de8\u6a21\u6001<\/td>\n<\/tr>\n \n \u7c7b\u522b\u7cfb\u7edf<\/td>\n \u56fa\u5b9a<\/td>\n \u56fa\u5b9a<\/td>\n \u56fa\u5b9a<\/td>\n \u5f00\u653e\u8bcd\u6c47<\/td>\n<\/tr>\n \n \u8f93\u5165<\/td>\n \u4ec5\u56fe\u50cf<\/td>\n \u4ec5\u56fe\u50cf<\/td>\n \u4ec5\u56fe\u50cf<\/td>\n \u56fe\u50cf + \u6587\u672c\u63d0\u793a<\/td>\n<\/tr>\n \n \u540e\u5904\u7406<\/td>\n NMS<\/td>\n \u65e0（\u5308\u7259\u5229\u5339\u914d）<\/td>\n \u65e0<\/td>\n \u65e0<\/td>\n<\/tr>\n \n \u65b0\u7c7b\u522b\u8fc1\u79fb<\/td>\n \u9700\u91cd\u8bad<\/td>\n \u9700\u91cd\u8bad<\/td>\n \u9700\u91cd\u8bad<\/td>\n \u6539\u63d0\u793a\u8bcd\u5373\u53ef，\u96f6\u6837\u672c<\/td>\n<\/tr>\n \n \u4ee3\u8868\u610f\u4e49<\/td>\n CNN \u5b9e\u65f6\u68c0\u6d4b\u5dc5\u5cf0<\/td>\n \u7aef\u5230\u7aef\u68c0\u6d4b\u5f00\u5c71<\/td>\n \u67b6\u6784\u901a\u7528\u6027\u8bc1\u660e<\/td>\n \u68c0\u6d4b\u8303\u5f0f\u9769\u65b0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u7ed3\u8bba：\u4ece YOLOv \u5230 DETR \/ YOLOS \u662f\u67b6\u6784\u5c42\u7684\u6f14\u8fdb（CNN \u2192 Transformer\u3001\u624b\u5de5\u7ec4\u4ef6 \u2192 \u7aef\u5230\u7aef），\u4ece\u5b83\u4eec\u5230 GLIP \u662f\u8303\u5f0f\u5c42\u7684\u8dc3\u8fc1（\u5c01\u95ed\u7c7b\u522b \u2192 \u5f00\u653e\u8bcd\u6c47\u3001\u89c6\u89c9\u72ec\u594f \u2192 \u8bed\u8a00\u5f15\u5bfc）\u3002\u56db\u4e2a\u6a21\u578b\u6784\u6210\u4e00\u6761\u5b8c\u6574\u7684\u6280\u672f\u8c31\u7cfb，\u4e5f\u662f\u7406\u89e3\u540e\u7eed\u5f00\u653e\u8bcd\u6c47\u5de5\u4f5c\u7684\u94a5\u5319\u3002<\/p>\n
\u516d\u3001\u8bfe\u7a0b\u8109\u7edc\u8854\u63a5<\/h4>\n
\u81f3\u6b64\u4e94\u4e2a\u6a21\u5757\u7684\u4e3b\u7ebf\u6e05\u6670\u53ef\u5faa：ViT（\u89c6\u89c9 Token \u5316）\u2192 CLIP（\u56fe\u50cf\u7ea7\u56fe\u6587\u5bf9\u9f50）\u2192 BLIP\/BLIP-2\/LLaVA（\u89c6\u89c9\u8fdb LLM \u7edf\u4e00\u63a8\u7406）\u2192 SAM（\u7c7b\u522b\u65e0\u5173\u5206\u5272）\u2192 GLIP（\u5bf9\u8c61\u7ea7\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b）\u3002GLIP \u4e0e SAM \u7684\u7ec4\u5408\u5373 Grounded-SAM（GLIP \u7528\u8bed\u8a00\u6307\u5b9a\u76ee\u6807\u5e76\u51fa\u6846 \u2192 SAM \u6cbf\u6846\u7cbe\u7ec6\u5206\u5272），\u662f\u5f53\u524d\u6700\u6d41\u884c\u7684\u5f00\u6e90\u5f00\u653e\u8bcd\u6c47\u5206\u5272\u6d41\u6c34\u7ebf\u4e4b\u4e00，\u4e5f\u4e3a\u540e\u7eed Stable Diffusion \u751f\u6210\u6a21\u5757（\u5c40\u90e8\u91cd\u7ed8\u9700\u5148\u5b9a\u4f4d\/\u5206\u5272）\u57cb\u4e0b\u4f0f\u7b14\u3002<\/p>\n
\u4e03\u3001\u5b66\u4e60\u8981\u70b9\u56de\u987e<\/h4>\n
- \n
- \u68c0\u6d4b Transformer \u5316\u4e24\u652f\u67f1：DETR \u7684 object query + \u5308\u7259\u5229\u5339\u914d；YOLOS \u7684\u88f8 ViT \u901a\u7528\u6027\u8bc1\u660e；<\/li>\n
- GLIP \u4e00\u53e5\u8bdd：\u68c0\u6d4b = "\u56fe\u50cf\u533a\u57df \u00d7 \u63d0\u793a\u77ed\u8bed"\u7684\u5bf9\u9f50\u95ee\u9898，\u8bcd\u6c47\u8868\u7531\u6587\u672c\u5f00\u653e；<\/li>\n
- \u5173\u952e\u67b6\u6784：DyHead + BERT + \u591a\u5c42\u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408（\u533a\u522b\u4e8e CLIP \u7684 late fusion）；<\/li>\n
- \u635f\u5931 = \u533a\u57df-\u8bcd token \u7ea7\u5bf9\u9f50\u5bf9\u6bd4\u635f\u5931 + (L1 + GIoU) \u5b9a\u4f4d\u635f\u5931，\u5308\u7259\u5229\u5339\u914d\u914d\u5bf9\u540e\u8054\u5408\u4f18\u5316；<\/li>\n
- \u6570\u636e\u914d\u65b9：3M \u4eba\u5de5\u6807\u6ce8 + 24M \u56fe\u6587\u5bf9\u81ea\u8bad\u7ec3\u4f2a\u6846；<\/li>\n
- \u52a8\u624b\u5efa\u8bae：Hugging Face \u76f4\u63a5\u52a0\u8f7d GLIP \u7cfb checkpoint，\u6539\u4e00\u884c\u63d0\u793a\u8bcd\u5373\u53ef\u4f53\u9a8c\u96f6\u6837\u672c\u68c0\u6d4b，\u518d\u8fdb\u9636\u7ec4\u5408 Grounded-SAM \u6d41\u6c34\u7ebf\u3002<\/li>\n<\/ul>\n
\n\u516b\u3001GLIP \u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218<\/h4>\n
\u524d\u9762\u51e0\u8282\u4ece\u539f\u7406\u4e0a\u7406\u89e3\u4e86 GLIP \u5982\u4f55\u628a\u76ee\u6807\u68c0\u6d4b\u91cd\u6784\u4e3a\u77ed\u8bed\u5b9a\u4f4d，\u672c\u8282\u7ed9\u51fa\u53ef\u76f4\u63a5\u8fd0\u884c\u7684\u5b9e\u6218\u4ee3\u7801，\u5e2e\u52a9\u8bfb\u8005\u5728 Hugging Face \u4e0a\u5feb\u901f\u4f53\u9a8c\u96f6\u6837\u672c\u68c0\u6d4b，\u5e76\u8fdb\u9636\u7ec4\u5408 Grounded-SAM \u5b8c\u6210\u4ece\u51fa\u6846\u5230\u5206\u5272\u7684\u5b8c\u6574\u843d\u5730\u3002<\/p>\n
1. \u4f7f\u7528 Hugging Face transformers \u52a0\u8f7d GLIP \u6a21\u578b<\/p>\n
transformers \u5e93\u63d0\u4f9b\u4e86 AutoModelForZeroShotObjectDetection \u63a5\u53e3，\u53ef\u7edf\u4e00\u52a0\u8f7d GLIP \u4e0e Grounding DINO \u7b49\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u6a21\u578b\u3002\u4e0b\u9762\u4ee5 GLIP \u4e3a\u4f8b\u7ed9\u51fa\u5b8c\u6574\u63a8\u7406\u6d41\u7a0b：<\/p>\n
from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
\nfrom PIL import Image
\nimport torch
\nimport matplotlib.pyplot as plt
\nimport matplotlib.patches as patches
\n1. \u52a0\u8f7d\u6a21\u578b\u4e0e\u5904\u7406\u5668（\u9996\u6b21\u8fd0\u884c\u4f1a\u81ea\u52a8\u4e0b\u8f7d checkpoint）
\nmodel_id = "microsoft\/grounding-dino-base" # \u4e5f\u53ef\u6362\u6210 GLIP \u7cfb checkpoint
\nprocessor = AutoProcessor.from_pretrained(model_id)
\nmodel = AutoModelForZeroShotObjectDetection.from_pretrained(model_id)
\nmodel.eval()
\n2. \u8bfb\u53d6\u56fe\u50cf\u5e76\u6784\u9020\u6587\u672c\u63d0\u793a（\u5404\u7c7b\u522b\u7528\u53e5\u53f7\u5206\u9694）
\nimage = Image.open("demo.jpg").convert("RGB")
\ntext = "cat . dog . person ."
\n3. \u8f93\u5165\u9884\u5904\u7406：\u56fe\u50cf\u7f29\u653e + \u6587\u672c token \u5316，\u7edf\u4e00\u9001\u5165\u6a21\u578b
\ninputs = processor(images=image, text=text, return_tensors="pt")
\n4. \u524d\u5411\u63a8\u7406，\u5f97\u5230\u9884\u6d4b\u6846\u3001\u7c7b\u522b\u4e0e\u7f6e\u4fe1\u5ea6
\nwith torch.no_grad():
\noutputs = model(**inputs)
\n5. \u9608\u503c\u8fc7\u6ee4 + \u6846\u5750\u6807\u8f6c\u6362（\u4ece\u6a21\u578b\u5750\u6807\u8fd8\u539f\u5230\u539f\u56fe\u50cf\u7d20\u5750\u6807）
\nresults = processor.post_process_grounded_object_detection(
\noutputs,
\ninputs.input_ids,
\nbox_threshold=0.35, # \u6846\u7f6e\u4fe1\u5ea6\u9608\u503c，\u4f4e\u4e8e\u6b64\u503c\u4e22\u5f03
\ntext_threshold=0.25, # \u6587\u672c\u5bf9\u9f50\u9608\u503c，\u63a7\u5236\u7c7b\u522b\u5339\u914d\u4e25\u683c\u5ea6
\ntarget_sizes=[image.size[::-1]] # \u8fd8\u539f\u5230\u539f\u56fe\u5c3a\u5bf8
\n)[0]
\n6. \u53ef\u89c6\u5316：\u5728\u539f\u56fe\u4e0a\u7ed8\u5236\u68c0\u6d4b\u6846\u4e0e\u6807\u7b7e
\nfig, ax = plt.subplots(1, 1, figsize=(12, 8))
\nax.imshow(image)
\nfor box, score, label in zip(results["boxes"], results["scores"], results["labels"]):
\nx1, y1, x2, y2 = box.tolist()
\nrect = patches.Rectangle((x1, y1), x2 – x1, y2 – y1,
\nlinewidth=2, edgecolor="red", facecolor="none")
\nax.add_patch(rect)
\nax.text(x1, y1 – 5, f"{label} {score:.2f}",
\ncolor="red", fontsize=12, bbox=dict(facecolor="white", alpha=0.8))
\nplt.axis("off")
\nplt.show() <\/p>\n\u5173\u952e\u6b65\u9aa4\u8bf4\u660e：\u7b2c 3 \u6b65\u7684 processor \u8d1f\u8d23\u628a\u56fe\u50cf\u7f29\u653e\u5230\u6a21\u578b\u8f93\u5165\u5c3a\u5bf8\u3001\u628a\u6587\u672c\u63d0\u793a\u7f16\u7801\u4e3a token \u5e8f\u5217；\u7b2c 5 \u6b65\u7684 box_threshold \u4e0e text_threshold \u5206\u522b\u63a7\u5236\u6846\u7f6e\u4fe1\u5ea6\u4e0e\u6587\u672c\u5bf9\u9f50\u7684\u8fc7\u6ee4\u5f3a\u5ea6，\u503c\u8d8a\u4f4e\u53ec\u56de\u8d8a\u591a\u3001\u8bef\u68c0\u4e5f\u8d8a\u591a；target_sizes \u7528\u4e8e\u628a\u6a21\u578b\u8f93\u51fa\u7684\u5f52\u4e00\u5316\u5750\u6807\u8fd8\u539f\u4e3a\u539f\u56fe\u50cf\u7d20\u5750\u6807，\u4fbf\u4e8e\u76f4\u63a5\u7ed8\u5236\u3002<\/p>\n
2. \u7ec4\u5408 Grounded-SAM：GLIP \u51fa\u6846 + SAM \u5206\u5272<\/p>\n
GLIP \u8d1f\u8d23\u7528\u8bed\u8a00\u6307\u5b9a\u76ee\u6807\u5e76\u8f93\u51fa\u8fb9\u754c\u6846，SAM \u518d\u6cbf\u6846\u5185\u533a\u57df\u505a\u7cbe\u7ec6\u5206\u5272，\u4e8c\u8005\u7ec4\u5408\u5373 Grounded-SAM \u6d41\u6c34\u7ebf\u3002\u7b80\u5316\u6d41\u7a0b\u5982\u4e0b：<\/p>\n
from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
\nfrom segment_anything import sam_model_registry, SamPredictor
\nfrom PIL import Image
\nimport numpy as np
\n\u7b2c\u4e00\u6b65：GLIP \u51fa\u6846（\u590d\u7528\u4e0a\u4e00\u8282\u7684\u52a0\u8f7d\u903b\u8f91）
\ndet_processor = AutoProcessor.from_pretrained("microsoft\/grounding-dino-base")
\ndet_model = AutoModelForZeroShotObjectDetection.from_pretrained("microsoft\/grounding-dino-base")
\nimage = Image.open("demo.jpg").convert("RGB")
\ntext = "cat . dog . person ."
\ninputs = det_processor(images=image, text=text, return_tensors="pt")
\noutputs = det_model(**inputs)
\nboxes = det_processor.post_process_grounded_object_detection(
\noutputs, inputs.input_ids,
\nbox_threshold=0.35, text_threshold=0.25,
\ntarget_sizes=[image.size[::-1]]
\n)[0]["boxes"].tolist()
\n\u7b2c\u4e8c\u6b65：SAM \u6cbf\u6846\u7cbe\u7ec6\u5206\u5272
\nsam = sam_model_registry"vit_h".to("cuda")
\npredictor = SamPredictor(sam)
\npredictor.set_image(np.array(image))
\nfor box in boxes:
\nx1, y1, x2, y2 = [int(v) for v in box]
\n# \u628a\u68c0\u6d4b\u6846\u4f5c\u4e3a SAM \u7684\u63d0\u793a\u6846，\u751f\u6210\u6846\u5185\u524d\u666f\u63a9\u7801
\nmasks, scores, _ = predictor.predict(
\nbox=np.array([x1, y1, x2, y2]),
\nmultimask_output=True
\n)
\nbest_mask = masks[scores.argmax()] # \u53d6\u7f6e\u4fe1\u5ea6\u6700\u9ad8\u7684\u63a9\u7801
\n# \u6b64\u5904\u53ef\u5c06 best_mask \u4fdd\u5b58\u4e3a PNG \u6216\u53e0\u52a0\u5230\u539f\u56fe\u7528\u4e8e\u540e\u7eed\u7f16\u8f91 <\/p>\n\u8fd9\u6bb5\u6d41\u7a0b\u628a GLIP \u7684\u5f00\u653e\u8bcd\u6c47\u5b9a\u4f4d\u80fd\u529b\u4e0e SAM \u7684\u7c7b\u522b\u65e0\u5173\u5206\u5272\u80fd\u529b\u4e32\u8054\u8d77\u6765：GLIP \u5148\u56de\u7b54"\u56fe\u4e2d\u54ea\u91cc\u6709 cat \/ dog \/ person"，SAM \u518d\u56de\u7b54"\u8fd9\u4e9b\u76ee\u6807\u5404\u81ea\u7684\u7cbe\u786e\u8f6e\u5ed3\u662f\u4ec0\u4e48"\u3002\u5728\u5f00\u653e\u8bcd\u6c47\u56fe\u50cf\u7f16\u8f91\u573a\u666f\u4e2d，\u8fd9\u4e00\u7ec4\u5408\u975e\u5e38\u5b9e\u7528\u2014\u2014\u4f8b\u5982\u7528\u6237\u8f93\u5165"\u628a\u56fe\u4e2d\u7684\u67f4\u72ac\u6362\u6210\u67ef\u57fa"，\u5373\u53ef\u5148\u7528 GLIP \u5b9a\u4f4d"\u67f4\u72ac"\u533a\u57df，\u518d\u7528 SAM \u5206\u5272\u51fa\u8be5\u533a\u57df\u63a9\u7801，\u6700\u540e\u4ea4\u7ed9 Stable Diffusion \u505a\u5c40\u90e8\u91cd\u7ed8，\u5b9e\u73b0\u76ee\u6807\u7ea7\u3001\u8bed\u4e49\u53ef\u63a7\u7684\u56fe\u50cf\u7f16\u8f91\u3002<\/p>\n
\u516b\u3001\u5b66\u4e60\u8def\u7ebf\u56fe<\/h4>\n
\u9488\u5bf9\u4e0d\u540c\u57fa\u7840\u7684\u5b66\u4e60\u8005，\u5efa\u8bae\u6309\u4ee5\u4e0b\u8def\u5f84\u5faa\u5e8f\u6e10\u8fdb\u5730\u638c\u63e1 GLIP \u53ca\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u6280\u672f：<\/p>\n
- \n
- \u5165\u95e8\u8def\u7ebf（0\u20132 \u5468）：\u5148\u638c\u63e1 Transformer \u57fa\u7840\u4e0e ViT \u539f\u7406，\u518d\u5b66\u4e60 DETR \u7684\u7aef\u5230\u7aef\u68c0\u6d4b\u8303\u5f0f，\u7406\u89e3 object query \u4e0e\u5308\u7259\u5229\u5339\u914d\u673a\u5236；<\/li>\n
- \u8fdb\u9636\u8def\u7ebf（2\u20134 \u5468）：\u6df1\u5165 YOLOS \u7684\u88f8 ViT \u68c0\u6d4b\u5b9e\u9a8c，\u7406\u89e3\u68c0\u6d4b\u80fd\u529b\u7684\u6d8c\u73b0\u673a\u5236；\u968f\u540e\u7cfb\u7edf\u5b66\u4e60 GLIP \u7684\u67b6\u6784\u4e09\u4ef6\u5957（DyHead + BERT + \u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408），\u638c\u63e1\u77ed\u8bed\u5b9a\u4f4d\u7684\u6838\u5fc3\u601d\u60f3；<\/li>\n
- \u5b9e\u6218\u8def\u7ebf（4\u20136 \u5468）：\u5728 Hugging Face \u4e0a\u52a0\u8f7d GLIP \u7cfb checkpoint，\u52a8\u624b\u5b9e\u8df5\u96f6\u6837\u672c\u68c0\u6d4b；\u8fdb\u9636\u7ec4\u5408 Grounded-SAM \u6d41\u6c34\u7ebf，\u5b8c\u6210\u4ece\u8bed\u8a00\u6307\u5b9a\u76ee\u6807\u5230\u7cbe\u7ec6\u5206\u5272\u7684\u5b8c\u6574\u843d\u5730；<\/li>\n
- \u7814\u7a76\u8def\u7ebf（6 \u5468\u4ee5\u4e0a）：\u7cbe\u8bfb GLIP \u8bba\u6587\u4e0e\u6e90\u7801，\u590d\u73b0\u635f\u5931\u51fd\u6570\u4e0e\u81ea\u8bad\u7ec3\u6570\u636e\u914d\u65b9，\u5c1d\u8bd5\u5728\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u4e0a\u5fae\u8c03，\u63a2\u7d22\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u7684\u8fb9\u754c\u4e0e\u6539\u8fdb\u65b9\u5411\u3002<\/li>\n<\/ul>\n
SEO \u5173\u952e\u8bcd：GLIP\u3001\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u3001\u591a\u6a21\u6001\u5927\u6a21\u578b\u3001phrase grounding\u3001YOLOS\u3001DETR \u539f\u7406\u3001DyHead\u3001\u96f6\u6837\u672c\u68c0\u6d4b\u3001GLIP \u635f\u5931\u51fd\u6570\u3001Grounded-SAM \u6559\u7a0b\u3001\u76ee\u6807\u68c0\u6d4b Transformer\u3001\u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408\u3002<\/p>\n
\u4e5d\u3001\u603b\u7ed3\u4e0e\u53c2\u8003\u8d44\u6599<\/h4>\n
\u672c\u6587\u7cfb\u7edf\u68b3\u7406\u4e86\u5362\u83c1\u535a\u58eb\u300a\u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b\u300b\u7b2c\u4e94\u6a21\u5757\u300aGLIP \u6a21\u578b\u8be6\u89e3\u300b（\u7b2c 16\u201319 \u96c6）\u7684\u6838\u5fc3\u5185\u5bb9：\u4ece DETR\u3001YOLOS \u7684\u68c0\u6d4b Transformer \u5316\u8bb2\u8d77，\u6df1\u5165\u5256\u6790 GLIP \u5982\u4f55\u628a\u76ee\u6807\u68c0\u6d4b\u91cd\u6784\u4e3a\u77ed\u8bed\u5b9a\u4f4d\u4efb\u52a1，\u5b9e\u73b0\u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u3002GLIP \u901a\u8fc7 DyHead + BERT + \u8bed\u8a00\u611f\u77e5\u6df1\u5ea6\u878d\u5408\u7684\u67b6\u6784\u8bbe\u8ba1，\u914d\u5408 27M \u9884\u8bad\u7ec3\u6570\u636e\u914d\u65b9\u4e0e\u8054\u5408\u635f\u5931\u51fd\u6570，\u5728\u96f6\u6837\u672c\u68c0\u6d4b\u4e0a\u53d6\u5f97\u4e86\u8d85\u8d8a\u6709\u76d1\u7763\u57fa\u7ebf\u7684\u6210\u7ee9，\u6807\u5fd7\u7740\u68c0\u6d4b\u8303\u5f0f\u4ece\u5c01\u95ed\u7c7b\u522b\u5230\u5f00\u653e\u8bcd\u6c47\u7684\u5173\u952e\u8dc3\u8fc1\u3002<\/p>\n
\u4ee5\u4e0b\u662f\u5b98\u65b9\u6587\u6863\u4e0e\u63a8\u8350\u9605\u8bfb\u8d44\u6599：<\/p>\n
- \n
- GLIP \u8bba\u6587：Grounded Language-Image Pre-training（CVPR 2022，\u5fae\u8f6f\u7814\u7a76\u9662）<\/li>\n
- Microsoft Research \u9879\u76ee\u9875：Object Detection in the Wild via Grounded Language-Image Pre-training<\/li>\n
- Hugging Face \u6a21\u578b\u5e93：\u641c\u7d22 GLIP \u5373\u53ef\u627e\u5230\u5b98\u65b9 checkpoint \u4e0e\u4f7f\u7528\u793a\u4f8b，\u652f\u6301\u96f6\u6837\u672c\u68c0\u6d4b\u5feb\u901f\u4f53\u9a8c<\/li>\n
- Grounded-SAM \u9879\u76ee：IDEA-Research\/Grounded-Segment-Anything，GLIP \u4e0e SAM \u7ec4\u5408\u7684\u5f00\u653e\u8bcd\u6c47\u5206\u5272\u6d41\u6c34\u7ebf<\/li>\n
- \u8bfe\u7a0b\u9009\u96c6：\u5362\u83c1\u535a\u58eb\u300a\u591a\u6a21\u6001\u5927\u6a21\u578b\u6559\u7a0b\u300b，\u7b2c 16\u201319 \u96c6\u5bf9\u5e94\u672c\u6a21\u5757\u5185\u5bb9<\/li>\n<\/ul>\n
<\/h3>\n
\u4e3b\u8981\u6765\u6e90：GLIP \u8bba\u6587 \u94fe\u63a5\u3001Microsoft Research \u9879\u76ee\u9875 \u94fe\u63a5\u3001GLIP \u6280\u672f\u89e3\u8bfb \u94fe\u63a5\u3001\u8bfe\u7a0b\u9009\u96c6 \u94fe\u63a5<\/p><\/p>\n","protected":false},"excerpt":{"rendered":"
\u76ee\u5f55：
\n\u4e00\u3001\u6a21\u5757\u5b9a\u4f4d：\u4ece\\”\u5206\u5272\u4e00\u5207\\”\u5230\\”\u68c0\u6d4b\u4e00\u5207\\”\u4e8c\u3001\u7b2c 16 \u96c6：YOLOS \u4e0e DETR\u2014\u2014\u68c0\u6d4b\u7684 Transformer \u5316\u4e09\u3001\u7b2c 17 \u96c6：GLIP \u539f\u7406\u2014\u2014\u68c0\u6d4b\u5373\u77ed\u8bed\u5b9a\u4f4d\u56db\u3001\u7b2c 18 \u96c6：GLIP \u635f\u5931\u51fd\u6570\u4e94\u3001\u7b2c 19 \u96c6：GLIP vs YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4\u516d\u3001\u8bfe\u7a0b\u8109\u7edc\u8854\u63a5\u4e03\u3001\u5b66\u4e60\u8981\u70b9\u56de\u987e\u516b\u3001\u5b66\u4e60\u8def\u7ebf\u56fe\u4e5d\u3001\u603b\u7ed3\u4e0e\u53c2\u8003\u8d44\u6599
\n\u6458\u8981：\u672c\u6587\u7cfb\u7edf\u68b3<\/p>\n","protected":false},"author":2,"featured_media":111726,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[11653,50,1665,523],"topic":[],"class_list":["post-111727","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-11653","tag-50","tag-1665","tag-523"],"yoast_head":"\nGLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/111727.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u76ee\u5f55： \u4e00\u3001\u6a21\u5757\u5b9a\u4f4d：\u4ece"\u5206\u5272\u4e00\u5207"\u5230"\u68c0\u6d4b\u4e00\u5207"\u4e8c\u3001\u7b2c 16 \u96c6：YOLOS \u4e0e DETR\u2014\u2014\u68c0\u6d4b\u7684 Transformer \u5316\u4e09\u3001\u7b2c 17 \u96c6：GLIP \u539f\u7406\u2014\u2014\u68c0\u6d4b\u5373\u77ed\u8bed\u5b9a\u4f4d\u56db\u3001\u7b2c 18 \u96c6：GLIP \u635f\u5931\u51fd\u6570\u4e94\u3001\u7b2c 19 \u96c6：GLIP vs YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4\u516d\u3001\u8bfe\u7a0b\u8109\u7edc\u8854\u63a5\u4e03\u3001\u5b66\u4e60\u8981\u70b9\u56de\u987e\u516b\u3001\u5b66\u4e60\u8def\u7ebf\u56fe\u4e5d\u3001\u603b\u7ed3\u4e0e\u53c2\u8003\u8d44\u6599 \u6458\u8981：\u672c\u6587\u7cfb\u7edf\u68b3\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/111727.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-01T08:50:23+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261001085022-6abe1ece5dac9.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/111727.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/111727.html\",\"name\":\"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-01T08:50:23+00:00\",\"dateModified\":\"2026-10-01T08:50:23+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/111727.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/111727.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/111727.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/111727.html","og_locale":"zh_CN","og_type":"article","og_title":"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u76ee\u5f55： \u4e00\u3001\u6a21\u5757\u5b9a\u4f4d：\u4ece\"\u5206\u5272\u4e00\u5207\"\u5230\"\u68c0\u6d4b\u4e00\u5207\"\u4e8c\u3001\u7b2c 16 \u96c6：YOLOS \u4e0e DETR\u2014\u2014\u68c0\u6d4b\u7684 Transformer \u5316\u4e09\u3001\u7b2c 17 \u96c6：GLIP \u539f\u7406\u2014\u2014\u68c0\u6d4b\u5373\u77ed\u8bed\u5b9a\u4f4d\u56db\u3001\u7b2c 18 \u96c6：GLIP \u635f\u5931\u51fd\u6570\u4e94\u3001\u7b2c 19 \u96c6：GLIP vs YOLOv \/ DETR \/ YOLOS \u5bf9\u6bd4\u516d\u3001\u8bfe\u7a0b\u8109\u7edc\u8854\u63a5\u4e03\u3001\u5b66\u4e60\u8981\u70b9\u56de\u987e\u516b\u3001\u5b66\u4e60\u8def\u7ebf\u56fe\u4e5d\u3001\u603b\u7ed3\u4e0e\u53c2\u8003\u8d44\u6599 \u6458\u8981：\u672c\u6587\u7cfb\u7edf\u68b3","og_url":"https:\/\/www.wsisp.com\/helps\/111727.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-01T08:50:23+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261001085022-6abe1ece5dac9.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/111727.html","url":"https:\/\/www.wsisp.com\/helps\/111727.html","name":"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-01T08:50:23+00:00","dateModified":"2026-10-01T08:50:23+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/111727.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/111727.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/111727.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"GLIP \u5f00\u653e\u8bcd\u6c47\u68c0\u6d4b\u8be6\u89e3\uff1a\u4ece DETR\/YOLOS \u5230\u77ed\u8bed\u5b9a\u4f4d\uff0c\u96f6\u6837\u672c\u68c0\u6d4b\u5b9e\u6218"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/111727","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=111727"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/111727\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/111726"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=111727"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=111727"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=111727"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=111727"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}