[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-327548-105":59,"doc-detail-327548-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","deltavid-enhancing-fine-grained-spatiotemporal-perception-with-cross-video-differences","DELTAVID - Enhancing Fine-Grained Spatiotemporal Perception with Cross-Video Differences","","Video multimodal large language models advance open-ended video understanding, yet often miss precise local spatiotemporal changes. When two videos share similar global semantics while differing only over a short time span or small region, existing methods provide unreliable evidence. DELTAVID introduces a verifiable proxy-task framework that uses cross-video spot-the-difference to train local change detection, temporal boundary judgment, and spatial evidence organization. Scalable datasets DELTAVID-10K and DELTAVID-Bench convert controllable real-video differences into labeled training and test samples, improving cross-video difference understanding and transferring local evidence abilities to broad video understanding benchmarks.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/deltavid-enhancing-fine-grained-spatiotemporal-perception-with-cross-video-differences/327548/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/deltavid-enhancing-fine-grained-spatiotemporal-perception-with-cross-video-differences/327548.png","ImageObject",300,407,{"name":92,"@type":93},"Violet","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-13","2026-09-21",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",16,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What problem does DELTAVID address in current video multimodal LLMs?","Question",{"text":113,"@type":114},"Current models may fail to detect small blurred regions, brief object appearances, or short temporal/spatial differences, especially when global semantics are similar, leading to unreliable local evidence.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How does DELTAVID turn cross-video differences into a trainable perception signal?",{"text":118,"@type":114},"DELTAVID models differences as spatial, temporal, and spatiotemporal local evidence so the model can identify local changes, judge temporal boundaries, and organize spatial evidence through comparisons of similar videos.",{"name":120,"@type":111,"acceptedAnswer":121},"What are DELTAVID-10K and DELTAVID-Bench used for?",{"text":122,"@type":114},"They scale verifiable evaluation and training by converting controllable local differences in real videos into evidence-labeled A/B samples, supporting both training supervision and benchmark testing.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},327548,1790468581,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":145,"read_time":146,"preview_page_count":4},4398048950312,"https://ap-avatar.wpscdn.com/avatar/400002538284de19e3c?_k=1778320343897328908","arXiv :2607 .02551v1 [ cs .CV] 26 Jun 2026  \nDELTAVID: Enhancing Fine-Grained Spatiotemporal  \nPerception  \nwith Cross-Video Differences  \nYankai Yang* 1,2 Yancheng Long* 1,2 Tingting Gao2 Han Li2  \nBin Wen† B 2 Shuo YangB 1  \nFan Yang2  \nAbstract  \nVideo multimodal large language models have made strong progress on open-ended video understanding, but they still lack precise local spatiotemporal perception.  \nWhen two videos share almost the same global semantics and differ only in a short time span or a small region, current models often fail to find the change and provide reliable evidence. We propose DELTAVID, a verifiable proxy-task framework that enhances fine-grained spatiotemporal perception with cross-video differences.  \nThe key idea is to turn cross-video spot-the-difference into a trainable perception signal, where a model identifies local changes, judges temporal boundaries, and organizes spatial evidence by comparing similar videos. To make this signal scalable to train and reliable to evaluate, we further introduce DELTAVID-10K and DELTAVID-Bench, which convert controllable local differences in real videos into evidence-labeled training and test samples. Experiments show that DELTAVID substantially improves performance on cross-video difference understanding and transfers the learned local evidence ability to general video understanding benchmarks, including MMVU, MLVU, Video-MME, VideoHolmes, VideoMMMU, LVBench, TempCompass, and LongVideoBench. These results show that crossvideo differences are not only an effective way to diagnose fine-grained perception failures, but also a scalable proxy supervision that moves Video MLLMs from coarse semantic understanding toward fine-grained spatiotemporal evidence reasoning.  \n1 Introduction  \nIn everyday visual perception, comparing two similar videos and identifying local changes is often intuitive: a human observer can align temporal progression, attend to local regions, and judge whether an object appears, moves, changes color, or changes speed. For video multimodal large language models (Video MLLMs) [1, 20, 27, 59, 2, 3, 56, 19, 6, 32], however, this capability does not emerge naturally. Although these models have made rapid progress in video question answering, long-video captioning, and open-ended reasoning, and can produce fluent global semantic descriptions, they may still miss a small blurred region, a short object appearance, a reversed or sped-up segment, ora local difference that lasts only a few seconds. This suggests that current video pretraining and instruction tuning paradigms can build high-level semantic alignment, but lack direct supervision for local changes, temporal boundaries, and spatial evidence.  \nReinforcement learning (RL) post-training has become an important way to improve multimodal models [14, 37, 36, 34, 33, 16, 50, 31, 55, 39] . Recent work shows that rule-based rewards can improve *Equal contribution. †Project leader. BCorresponding authors.  \n1Harbin Institute of Technology, Shenzhen. 2 Kuaishou Technology.  \nCorrespondence to: Shuo Yang \u003C[shuoyang@hit.edu.cn](shuoyang@hit.edu.cn)>, Bin Wen \u003C[wenbin@kuaishou.com](wenbin@kuaishou.com)>.  \nDeltaVid Data Construction Pipeline  \nReal Video Clip  \nClip Sampling (3~60s)  \n1) Spatial Perturbations  \nbbox shape mosaic color shift  \n2) Temporal Perturbations  \n3) Spatial-temporal Perturbations  \nspeed up  \nreverse  \nfreeze  \ndelete  \nloop  \nappear disappear moving  \nGroup Relative Policy Optimization  \nVideo MLLM  \nGroup Computation  \n| Anti-shortcut design |  |\n| --- | --- |\n| | Noise Injection (visual, temporal, compression)\u003Cbr>Randomization (order, types, positions, colors)\u003Cbr>Overlap Control (limit co-occurence of cues) |\n\nFigure 1: DELTAVID data construction, rule-reward training, and anti-shortcut design.  \nreasoning format, answer quality, and some perception skills without a human preference model. Video methods have also started to explore temporal reasoning and video decision making [","cbCaihR7fypR1rr4","https://ap.wps.com/l/cbCaihR7fypR1rr4","pdf",18487422,21,"English","# Introduction\n## Fine-grained spatiotemporal perception via cross-video differences\n## DELTAVID framework and verifiable proxy task\n## DELTAVID-10K and DELTAVID-Bench datasets\n## Experiments and transfer to video understanding benchmarks","[{\"question\":\"What problem does DELTAVID address in current video multimodal LLMs?\",\"answer\":\"Current models may fail to detect small blurred regions, brief object appearances, or short temporal/spatial differences, especially when global semantics are similar, leading to unreliable local evidence.\"},{\"question\":\"How does DELTAVID turn cross-video differences into a trainable perception signal?\",\"answer\":\"DELTAVID models differences as spatial, temporal, and spatiotemporal local evidence so the model can identify local changes, judge temporal boundaries, and organize spatial evidence through comparisons of similar videos.\"},{\"question\":\"What are DELTAVID-10K and DELTAVID-Bench used for?\",\"answer\":\"They scale verifiable evaluation and training by converting controllable local differences in real videos into evidence-labeled A/B samples, supporting both training supervision and benchmark testing.\"}]","DELTAVID - Enhancing Fine-Grained Spatiotemporal Perception with Cross-Video Differences | PDF",1789977492,53]