
{"id":10608,"date":"2026-08-01T03:16:54","date_gmt":"2026-07-31T19:16:54","guid":{"rendered":"https:\/\/infernews.com\/blog\/a-method-for-multi-reference-image-grounded-video-captioning\/"},"modified":"2026-08-01T03:18:21","modified_gmt":"2026-07-31T19:18:21","slug":"a-method-for-multi-reference-image-grounded-video-captioning","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/a-method-for-multi-reference-image-grounded-video-captioning\/","title":{"rendered":"RefCaptioner\uff1a\u53c3\u8003\u5716\u7d81\u5b9a\u5c0d\u61c9\u5f71\u7247\u5b57\u5e55"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-f6467a4d2036.jpg\" alt=\"RefCaptioner grounds local caption phrases to relevant reference images while rejecting distractors.\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u505a\u5f71\u7247\u63cf\u8ff0\u6642\uff0c\u6700\u6613\u51fa\u932f\u5514\u4fc2\u53e5\u5b50\u5beb\u5f97\u5514\u5920\u9577\uff0c\u800c\u4fc2\u8b1b\u5230\u67d0\u500b\u4eba\u3001\u7269\u4ef6\u6216\u89d2\u5ea6\u6642\uff0c\u7121\u6cd5\u4ea4\u4ee3\u6587\u5b57\u7a76\u7adf\u5c0d\u61c9\u908a\u5f35\u53c3\u8003\u5716\u3002RefCaptioner\u5c6c\u65bc\u5f71\u7247\u5b57\u5e55\u751f\u6210\u6a21\u578b\u9805\u76ee\uff0c\u96c6\u4e2d\u8655\u7406 multi-reference image-grounded video captioning\uff1a\u4e00\u908a\u4fdd\u7559\u7d30\u7bc0\u8207\u4e8b\u5be6\u6e96\u78ba\u5ea6\uff0c\u4e00\u908a\u5c07\u5c40\u90e8\u63cf\u8ff0\u540c\u5019\u9078\u53c3\u8003\u5716\u660e\u78ba\u7d81\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">RefCaptioner \u4e0d\u53ea\u662f\u628a\u6240\u6709\u53c3\u8003\u5716\u585e\u5165\u8f38\u51fa\uff0c\u800c\u4fc2\u6703\u6311\u9078\u771f\u6b63\u6709\u7528\u7684\u5716\uff0c\u5c07\u5c0d\u61c9\u7247\u8a9e\u52a0\u4e0a <code>&lt;Image_N&gt;<\/code> \u6a19\u7c64\uff0c\u9047\u5230\u540c\u4e00\u4e3b\u9ad4\u7684\u4e0d\u540c\u8996\u89d2\u53c8\u6703\u505a\u5206\u7d44\uff0c\u5f71\u7247\u6839\u672c\u7121\u51fa\u73fe\u7684\u5167\u5bb9\u5c31\u76f4\u63a5\u7565\u904e\u3002\u5462\u7a2e\u8a2d\u8a08\u6e1b\u5c11\u932f\u914d\u540c\u8aa4\u5c0e\uff0c\u6bd4\u8d77\u53ea\u8ffd\u6c42\u6d41\u66a2\u5b57\u5e55\uff0c\u66f4\u8457\u91cd\u53ef\u6838\u5c0d\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6280\u8853\u4e0a\uff0cRefCaptioner\u7528\u5169\u6bb5\u5f0f post-training\u3002\u5148\u4ee5 capability-preserving SFT \u5b78\u6703 grounded caption \u683c\u5f0f\uff0c\u540c\u6642\u76e1\u91cf\u4fdd\u7559\u4e00\u822c captioning \u80fd\u529b\uff1b\u4e4b\u5f8c\u518d\u7528 Hierarchical Coverage-Discounted GRPO\uff08HCD-GRPO\uff09\u540c\u6642\u512a\u5316 factual-caption \u5206\u652f\u8207 multi-reference grounding \u5206\u652f\uff0c\u4e26\u52a0\u5165 deterministic guards\uff0c\u907f\u514d\u7522\u751f\u683c\u5f0f\u932f\u8aa4\u6216\u6307\u5411\u4e0d\u5b58\u5728\u5716\u7247\u7684\u6a19\u7c64\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u63d0\u4f9b\u5b98\u65b9 inference pipeline\u3001SFT \u8cc7\u6599\u6e96\u5099\u3001HCD-GRPO \u8a13\u7df4\u540c MRVBench evaluation pipeline<\/li>\n<li>\u5df2\u516c\u958b\u8ad6\u6587\u8207\u6a21\u578b\u6b0a\u91cd\uff0c\u4ea6\u6709 Data Format\u3001Training\u3001Evaluation \u6587\u4ef6\u53ef\u8ddf\u9032<\/li>\n<li>\u74b0\u5883\u5206\u6210\u4e3b\u74b0\u5883\u8207 GRPO \u5c08\u7528 veRL\/vLLM \u74b0\u5883\uff0c\u4ee3\u8868\u8a13\u7df4\u6d41\u7a0b\u8f03\u5b8c\u6574\u4f46\u914d\u7f6e\u4ea6\u8f03\u8b1b\u7a76<\/li>\n<li>\u9069\u5408\u505a\u5f71\u7247\u7406\u89e3\u3001\u8cc7\u6599\u6a19\u8a3b\u3001\u591a\u93e1\u982d\u4eba\u7269\u6558\u8ff0\u540c\u9700\u8981\u6aa2\u67e5\u5716\u6587\u5c0d\u61c9\u7684\u7814\u7a76\u5718\u968a<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u90e8\u7f72\u8207\u6e2c\u8a66\uff1a\u63a8\u7406\u3001SFT\u3001\u8a55\u4f30\u5171\u7528\u4e3b\u74b0\u5883\uff0cGRPO \u53e6\u8a2d\u4e00\u5957\u74b0\u5883\uff0c\u4e26\u4e14\u8981\u5c0d\u6307\u5b9a veRL \u7248\u672c\u5957\u7528 patch\uff0c\u53cd\u6620\u4f62\u8f03\u504f\u7814\u7a76\u578b\u5de5\u4f5c\u6d41\uff0c\u800c\u5514\u4fc2\u4e0b\u8f09\u5373\u7528\u7684\u5c0f\u5de5\u5177\u3002\u6548\u80fd\u6578\u5b57\u5728\u63d0\u4f9b\u7684\u5167\u5bb9\u672a\u898b\u5b8c\u6574\u5c55\u958b\uff0c\u4f46\u65e2\u7136\u5df2\u9644 MRVBench evaluation pipeline\uff0c\u81f3\u5c11\u8868\u793a\u4f5c\u8005\u6709\u628a\u300c\u5b57\u5e55\u5beb\u5f97\u5c0d\u300d\u540c\u300c\u5716\u6587\u5c0d\u5f97\u6e96\u300d\u5206\u958b\u6aa2\u9a57\uff0c\u8f03\u9069\u5408\u91cd\u8996\u53ef\u89e3\u91cb\u8f38\u51fa\u7684\u5718\u968a\u63a1\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/pkucs-Ltf\/RefCaptioner\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/TengfeiLiuCoder\/RefCaptioner\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>RefCaptioner\u5514\u53ea\u5beb\u5f71\u7247\u5b57\u5e55\uff0c\u4ef2\u6703\u6e05\u695a\u6a19\u793a\u53e5\u4e2d\u63cf\u8ff0\u5c0d\u61c9\u908a\u5f35\u53c3\u8003\u5716\u3002\u5c0d\u8981\u8655\u7406\u591a\u8996\u89d2\u4eba\u7269\u6216\u6df7\u5165\u5e72\u64fe\u5716\u7684\u60c5\u5883\uff0c\u5462\u7a2e\u505a\u6cd5\u5be6\u7528\u5f97\u591a\u3002<\/p>\n","protected":false},"author":8,"featured_media":10607,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,116,31,30,157,76,127],"tags":[],"class_list":["post-10608","post","type-post","status-publish","format-standard","hentry","category-133","category-agentic","category-video","category-image","category-157","category-76","category-127"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10608","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=10608"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10608\/revisions"}],"predecessor-version":[{"id":10611,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10608\/revisions\/10611"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10607"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10608"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10608"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10608"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}