{"id":3515,"date":"2025-06-16T13:51:11","date_gmt":"2025-06-16T05:51:11","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=3515"},"modified":"2025-06-16T13:51:11","modified_gmt":"2025-06-16T05:51:11","slug":"202506-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-grounding-dino-marrying-dino-with-grounded-pre-training-for-open-set-object-detection","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202506-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-grounding-dino-marrying-dino-with-grounded-pre-training-for-open-set-object-detection\/","title":{"rendered":"202506 \u8bba\u6587\u7814\u8bfb-Grounding DINO: Marrying DINO with Grounded Pre-training for Open-Set Object Detection"},"content":{"rendered":"\n

\u6765\u6e90\uff1aECCV 2024<\/p>\n\n\n\n

\u4f5c\u8005\uff1aShilong Liu\u3001Zhaoyang Zeng\u3001Tianhe Ren\u3001Feng Li\u3001Hao Zhang…<\/p>\n\n\n\n

\u5355\u4f4d\uff1aIDEA <\/p>\n\n\n\n

\u4e00\u3001\u80cc\u666f<\/strong><\/p>\n\n\n\n

\u4f20\u7edf\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u5982 Faster R-CNN \u6216 DINO\uff0c\u4e3b\u8981\u662f \u95ed\u96c6\u76ee\u6807\u68c0\u6d4b<\/strong>\uff08Closed-set Object Detection\uff09\uff1a\u53ea\u80fd\u8bc6\u522b\u8bad\u7ec3\u65f6\u89c1\u8fc7\u7684\u56fa\u5b9a\u7c7b\u522b\u3002\u800c\u73b0\u5b9e\u4e16\u754c\u4e2d\u5b58\u5728\u5927\u91cf\u201c\u5f00\u653e\u7c7b\u522b\u201d\uff0c\u56e0\u6b64\u9700\u8981\u80fd\u5904\u7406 \u5f00\u653e\u96c6\u76ee\u6807\u68c0\u6d4b<\/strong>\uff08Open-set Object Detection, OOD\uff09\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

\u56e0\u6b64\uff0c\u672c\u6587\u76ee\u6807\u662f\u6784\u5efa\u4e00\u4e2a\u53ef\u901a\u8fc7\u4eba\u7c7b\u8bed\u8a00\u8f93\u5165\u68c0\u6d4b\u4efb\u610f\u7269\u4f53<\/strong>\u7684\u5f00\u653e\u96c6\u76ee\u6807\u68c0\u6d4b\u6a21\u578b \u2014\u2014 Grounding DINO\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u8d21\u732e<\/strong><\/p>\n\n\n\n

    \n
  1. \u63d0\u51fa Grounding DINO \u6a21\u578b<\/strong>\uff1a\u7ed3\u5408 DINO \u7684 Transformer \u68c0\u6d4b\u67b6\u6784\u4e0e grounding \u9884\u8bad\u7ec3\uff0c\u652f\u6301\u4ee5\u81ea\u7136\u8bed\u8a00\u68c0\u6d4b\u4efb\u610f\u7269\u4f53\u3002<\/li>\n\n\n\n
  2. \u4e09\u9636\u6bb5\u6a21\u6001\u878d\u5408\u673a\u5236<\/strong>\uff1a\u5728Neck\u3001Query \u521d\u59cb\u5316\u4e0e Decoder \u4e2d\u5747\u5b9e\u73b0\u8bed\u8a00\u4e0e\u56fe\u50cf\u7279\u5f81\u878d\u5408\uff0c\u5373\u7d27\u8026\u5408\u6a21\u6001\u878d\u5408\uff08tight modality fusion\uff09\u3002<\/li>\n\n\n\n
  3. \u5b50\u53e5\u7ea7\u6587\u672c\u8868\u793a<\/strong>\uff1a\u63d0\u51faSub-sentence\u6587\u672c\u8868\u793a\u6cd5\uff0c\u901a\u8fc7 attention mask \u5c4f\u853d\u65e0\u5173\u7c7b\u522b\u95f4\u5e72\u6270\uff0c\u63d0\u5347\u8bed\u8a00\u7279\u5f81\u8868\u8fbe\u80fd\u529b\u3002<\/li>\n\n\n\n
  4. \u652f\u6301\u6307\u4ee3\u8868\u8fbe\u68c0\u6d4b\u4efb\u52a1\uff08REC\uff09<\/strong>\uff1a\u6269\u5c55\u5f00\u653e\u96c6\u68c0\u6d4b\u6a21\u578b\u7528\u4e8e RefCOCO\/+\/g \u7b49\u4efb\u52a1\u3002<\/li>\n<\/ol>\n\n\n\n

    \u4e09\u3001\u65b9\u6cd5<\/strong><\/p>\n\n\n\n

      \n
    1. \u6574\u4f53\u67b6\u6784<\/strong><\/li>\n<\/ol>\n\n\n\n
        <\/ol>\n\n\n\n

        Grounding DINO \u91c7\u7528\u53cc\u7f16\u7801\u5668 + \u5355\u89e3\u7801\u5668\u7684\u7ed3\u6784\uff1a\u56fe\u50cf\u7f16\u7801\u5668\u4f7f\u7528 Swin Transformer \u63d0\u53d6\u591a\u5c3a\u5ea6\u56fe\u50cf\u7279\u5f81\uff0c\u6587\u672c\u7f16\u7801\u5668\u4f7f\u7528 BERT \u63d0\u53d6\u6587\u672c token \u5411\u91cf\u3002\u968f\u540e\u901a\u8fc7\u7279\u5f81\u589e\u5f3a\u5668\u8fdb\u884c\u591a\u6a21\u6001\u878d\u5408\uff0c\u518d\u4f7f\u7528\u8bed\u8a00\u5f15\u5bfc\u7684 query \u9009\u62e9\u6a21\u5757\u751f\u6210\u68c0\u6d4b query\uff0c\u6700\u540e\u7ecf\u7531\u8de8\u6a21\u6001\u89e3\u7801\u5668\u8f93\u51fa\u76ee\u6807\u6846\u53ca\u5bf9\u5e94\u6587\u672c\u3002<\/p>\n\n\n\n

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

        2. \u7279\u5f81\u589e\u5f3a\u5668\uff08Feature Enhancer\uff09<\/strong><\/p>\n\n\n\n

        \u8be5\u6a21\u5757\u5bf9\u56fe\u50cf\u7279\u5f81\u548c\u6587\u672c\u7279\u5f81\u8fdb\u884c\u4ea4\u53c9\u878d\u5408\uff0c\u7531\u591a\u5c42\u5806\u53e0\u7ec4\u6210\uff1a<\/p>\n\n\n\n

          \n
        • \u56fe\u50cf\u901a\u9053\uff1a\u4f7f\u7528 deformable self-attention \u63d0\u5347\u56fe\u50cf\u7a7a\u95f4\u5efa\u6a21\u80fd\u529b\uff1b<\/li>\n\n\n\n
        • \u6587\u672c\u901a\u9053\uff1a\u4f7f\u7528 vanilla self-attention \u63d0\u53d6\u4e0a\u4e0b\u6587\u8bed\u4e49\uff1b<\/li>\n\n\n\n
        • \u56fe\u6587\u878d\u5408\uff1a\u901a\u8fc7 image-to-text \u548c text-to-image cross-attention \u5b9e\u73b0\u8de8\u6a21\u6001\u5bf9\u9f50\uff0c\u589e\u5f3a\u5c40\u90e8\u4e0e\u5168\u5c40\u8bed\u4e49\u7406\u89e3\u3002<\/li>\n<\/ul>\n\n\n\n
          \"\"<\/figure>\n\n\n\n

          3. \u8bed\u8a00\u5f15\u5bfc\u7684 Query \u9009\u62e9\uff08Language-Guided Query Selection\uff09<\/strong><\/p>\n\n\n\n

          \u8be5\u6a21\u5757\u4ece\u56fe\u50cf\u7f16\u7801\u7279\u5f81\u4e2d\u9009\u51fa\u4e0e\u6587\u672c\u6700\u76f8\u5173\u7684\u82e5\u5e72\u533a\u57df\u4f5c\u4e3a decoder \u8f93\u5165\u7684\u521d\u59cb query\uff1a<\/p>\n\n\n\n

            \n
          • \u9996\u5148\u8ba1\u7b97\u56fe\u50cf\u7279\u5f81\u4e0e\u6587\u672c\u7279\u5f81\u7684\u76f8\u4f3c\u5ea6\u77e9\u9635\uff1b<\/li>\n\n\n\n
          • \u5bf9\u6bcf\u4e2a\u56fe\u50cf\u4f4d\u7f6e\u53d6\u5176\u4e0e\u6240\u6709\u6587\u672c token \u7684\u6700\u5927\u76f8\u4f3c\u5ea6\uff1b<\/li>\n\n\n\n
          • \u6309\u5f97\u5206\u9009\u51fa\u524d Nq\uff08\u5982 900\uff09\u4e2a\u4f4d\u7f6e\uff1b<\/li>\n\n\n\n
          • \u751f\u6210\u5bf9\u5e94\u7684 content \u548c positional query\uff08\u5176\u4e2d\u4f4d\u7f6e\u91c7\u7528\u52a8\u6001 anchor box \u521d\u59cb\u5316\uff0c\u5185\u5bb9\u90e8\u5206\u4e3a\u53ef\u5b66\u4e60\u53c2\u6570\uff09\u3002<\/li>\n<\/ul>\n\n\n\n

            4. \u8de8\u6a21\u6001\u89e3\u7801\u5668\uff08Cross-Modality Decoder\uff09<\/strong><\/p>\n\n\n\n

            \u6bcf\u5c42\u89e3\u7801\u5668\u5305\u542b\u56db\u90e8\u5206\uff1a<\/p>\n\n\n\n

              \n
            • \u81ea\u6ce8\u610f\u529b\u5c42\uff08Self-Attention\uff09\uff1b<\/li>\n\n\n\n
            • \u56fe\u50cf cross-attention\uff08\u4ece\u56fe\u50cf\u7279\u5f81\u4e2d\u63d0\u53d6\u76f8\u5173\u533a\u57df\uff09\uff1b<\/li>\n\n\n\n
            • \u6587\u672c cross-attention\uff08\u589e\u5f3a\u8bed\u8a00\u8bed\u4e49\u5bf9 query \u7684\u6307\u5bfc\uff09\uff1b<\/li>\n\n\n\n
            • FFN\uff08\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff09\u66f4\u65b0 query \u8868\u5f81\u3002<\/li>\n<\/ul>\n\n\n\n

              \u4e0e DINO \u76f8\u6bd4\uff0c\u8be5\u89e3\u7801\u5668\u5f15\u5165\u989d\u5916\u7684\u6587\u672c cross-attention\uff0c\u4f7f\u6bcf\u5c42\u90fd\u6ce8\u5165\u8bed\u8a00\u4fe1\u606f\u3002<\/p>\n\n\n\n

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

              5. \u5b50\u53e5\u7ea7\u6587\u672c\u8868\u793a<\/strong><\/p>\n\n\n\n

              \u9488\u5bf9\u591a\u4e2a\u7c7b\u522b\u540d\u62fc\u63a5\u4e3a\u53e5\u5b50\u7684\u60c5\u51b5\uff0c\u8be5\u65b9\u6cd5\u5f15\u5165 attention mask\uff0c\u963b\u6b62\u4e0d\u540c\u7c7b\u522b\u8bcd\u4e4b\u95f4\u4ea7\u751f\u6ce8\u610f\u529b\u8054\u7cfb\uff1a<\/p>\n\n\n\n

                \n
              • \u907f\u514d\u65e0\u5173\u8bcd\u4e4b\u95f4\u4e92\u76f8\u5e72\u6270\uff1b<\/li>\n\n\n\n
              • \u540c\u65f6\u4fdd\u7559\u6bcf\u4e2a\u8bcd\u7684\u72ec\u7acb\u5411\u91cf\uff0c\u63d0\u5347\u7ec6\u7c92\u5ea6\u7684\u8bed\u4e49\u8868\u793a\u80fd\u529b\uff1b<\/li>\n\n\n\n
              • \u5728 grounding \u8bad\u7ec3\u4e2d\u66f4\u7b26\u5408\u68c0\u6d4b\u7c7b token \u7684\u4f7f\u7528\u573a\u666f\u3002<\/li>\n<\/ul>\n\n\n\n

                6. \u635f\u5931\u51fd\u6570\u8bbe\u8ba1<\/strong><\/p>\n\n\n\n

                \u91c7\u7528 DETR \u7cfb\u5217\u7684\u6807\u51c6\u635f\u5931\u7ec4\u5408\uff1a<\/p>\n\n\n\n

                  \n
                • \u8fb9\u754c\u6846\u56de\u5f52<\/strong>\uff1aL1 + GIOU Loss\uff1b<\/li>\n\n\n\n
                • \u76ee\u6807\u5206\u7c7b<\/strong>\uff1aquery \u4e0e\u6240\u6709\u6587\u672c token \u505a\u70b9\u79ef\u5f97\u5230\u9884\u6d4b logits\uff0c\u4f7f\u7528 Focal Loss \u8fdb\u884c supervision\uff1b<\/li>\n\n\n\n
                • \u4f7f\u7528 Hungarian \u5339\u914d\uff08\u57fa\u4e8e\u5206\u7c7b\u635f\u5931 + \u56de\u5f52\u635f\u5931\uff09\u51b3\u5b9a\u6b63\u8d1f\u6837\u672c\u5bf9\u9f50\u5173\u7cfb\uff1b<\/li>\n\n\n\n
                • \u6bcf\u5c42 decoder \u548c encoder \u8f93\u51fa\u5747\u52a0\u5165\u8f85\u52a9\u635f\u5931\uff08auxiliary loss\uff09\u4ee5\u7a33\u5b9a\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n\n\n\n

                  \u56db\u3001\u5b9e\u9a8c<\/strong><\/p>\n\n\n\n

                    \n
                  1. Zero-shot Transfer \u96f6\u6837\u672c\u8fc1\u79fb<\/li>\n<\/ol>\n\n\n\n
                      <\/ol>\n\n\n\n
                        \n
                      • COCO<\/strong>\uff1aZero-shot \u4e0b\u8fbe\u6210 52.5 AP\uff0c\u4f18\u4e8e DINO \u548c GLIP<\/li>\n<\/ul>\n\n\n\n
                        \"\"<\/figure>\n\n\n\n
                          \n
                        • LVIS<\/strong>\uff1a\u5728\u5e38\u89c1\u7c7b\u522b\u8868\u73b0\u4f18\u4e8e GLIP\uff0c\u4f46\u5728\u7a00\u6709\u7c7b\u522b\u7565\u900a<\/li>\n<\/ul>\n\n\n\n
                          \"\"<\/figure>\n\n\n\n
                            \n
                          • ODinW<\/strong>\uff1a\u5728 35 \u4e2a\u771f\u5b9e\u4e16\u754c\u6570\u636e\u96c6\u4e2d\u5e73\u5747 AP \u8fbe 26.1\uff0c\u8d85\u8fc7 Florence \u7b49\u5927\u578b\u6a21\u578b<\/li>\n<\/ul>\n\n\n\n
                            \"\"<\/figure>\n\n\n\n

                            2. Referring Expression Comprehension<\/p>\n\n\n\n

                            \u672a\u5fae\u8c03\u65f6\u6548\u679c\u4e00\u822c\uff1b\u52a0\u5165 RefCOCO \u8bad\u7ec3\u540e\u5927\u5e45\u63d0\u5347\uff0c\u63d0\u793a\u5f53\u524d\u6a21\u578b\u5728\u7ec6\u7c92\u5ea6\u68c0\u6d4b\u4efb\u52a1\u4ecd\u6709\u53d1\u5c55\u7a7a\u95f4<\/p>\n\n\n\n

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

                            \u4e94\u3001\u603b\u7ed3<\/strong><\/p>\n\n\n\n

                            Grounding DINO \u662f\u5f00\u653e\u96c6\u68c0\u6d4b\u9886\u57df\u7684\u91cd\u8981\u8fdb\u5c55\u3002\u901a\u8fc7\u5bf9 Transformer \u68c0\u6d4b\u5668\u7684\u591a\u6a21\u6001\u6269\u5c55\u4e0e\u8bed\u8a00\u611f\u77e5\u673a\u5236\u8bbe\u8ba1\uff0c\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u901a\u7528\u6027\u548c\u6269\u5c55\u6027\u3002\u5728\u56fe\u6587\u7ed3\u5408\u68c0\u6d4b\u3001\u6307\u4ee3\u7406\u89e3\u7b49\u4efb\u52a1\u4e2d\u5177\u5907\u9886\u5148\u6027\u80fd\u3002\u672a\u6765\u5de5\u4f5c\u53ef\u8fdb\u4e00\u6b65\u62d3\u5c55\u5230\u5b9e\u4f8b\u5206\u5272\u3001\u591a\u6a21\u6001\u7f16\u8f91\u3001\u590d\u6742\u89c6\u89c9\u95ee\u7b54\u7b49\u573a\u666f\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

                            \u6765\u6e90\uff1aECCV 2024 \u4f5c\u8005\uff1aShilong Liu\u3001Zhaoyang Zeng\u3001Tianhe Ren\u3001Fe […]<\/p>\n","protected":false},"author":64,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3515","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3515","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/64"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=3515"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3515\/revisions"}],"predecessor-version":[{"id":3570,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3515\/revisions\/3570"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=3515"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=3515"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=3515"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}