Hugging Face Daily Papers · August 31, 2026 · 4 min read

Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

ABot-Recon turns a single continuous video into a globally consistent 3D reconstruction in real time. Whether walking around a building with a phone, driving through city streets with a dashcam, or flying a drone over a campus, it reconstructs long video streams using only a fixed 12-frame local context, composing current-frame geometry and adjacent relative poses without persistent learned long-range memory. It scales to 10,000 + frame sequences at 24.45 FPS with only 6.71 GB of GPU memory, enabling efficient real-time 3D reconstruction of large-scale environments.\n<a href=\"https://cdn-uploads.huggingface.co/production/uploads/6433bc32a4c9c55871a42812/Ox12Ux8aOcghXL9_tCdFg.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/6433bc32a4c9c55871a42812/Ox12Ux8aOcghXL9_tCdFg.png\" alt=\"teaser\"></a>\n","updatedAt":"2026-08-31T02:46:48.551Z","author":{"_id":"6433bc32a4c9c55871a42812","avatarUrl":"/avatars/c27c063d02419b8ba32657514c07e3b2.svg","fullname":"qian#143","name":"qian43","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":7,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7899578809738159},"editors":["qian43"],"editorAvatarUrls":["/avatars/c27c063d02419b8ba32657514c07e3b2.svg"],"reactions":[{"reaction":"🚀","users":["qian43"],"count":1}],"isReport":false}},{"id":"6a94f815b10e7f5eae975e61","author":{"_id":"6433bc32a4c9c55871a42812","avatarUrl":"/avatars/c27c063d02419b8ba32657514c07e3b2.svg","fullname":"qian#143","name":"qian43","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":7,"isUserFollowing":false},"createdAt":"2026-08-31T03:42:13.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"https://www.youtube.com/watch?v=06OM-4oCw3s","html":"<a href=\"https://www.youtube.com/watch?v=06OM-4oCw3s\" rel=\"nofollow\">https://www.youtube.com/watch?v=06OM-4oCw3s</a>\n","updatedAt":"2026-08-31T03:42:13.327Z","author":{"_id":"6433bc32a4c9c55871a42812","avatarUrl":"/avatars/c27c063d02419b8ba32657514c07e3b2.svg","fullname":"qian#143","name":"qian43","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":7,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.4163947105407715},"editors":["qian43"],"editorAvatarUrls":["/avatars/c27c063d02419b8ba32657514c07e3b2.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.27529","authors":[{"_id":"6a94e198073195fee51571fd","name":"Jiarong Han","hidden":false},{"_id":"6a94e198073195fee51571fe","name":"Jincheng Xiong","hidden":false},{"_id":"6a94e198073195fee51571ff","name":"Yuzhou Liu","hidden":false},{"_id":"6a94e198073195fee5157200","name":"Linzhe Shi","hidden":false},{"_id":"6a94e198073195fee5157201","name":"Changjie Wu","hidden":false},{"_id":"6a94e198073195fee5157202","name":"Ning Guo","hidden":false},{"_id":"6a94e198073195fee5157203","name":"Mu Xu","hidden":false},{"_id":"6a94e198073195fee5157204","name":"Hang Zhang","hidden":false},{"_id":"6a94e198073195fee5157205","name":"Ming Qian","hidden":false}],"publishedAt":"2026-08-27T00:00:00.000Z","submittedOnDailyAt":"2026-08-31T00:00:00.000Z","title":"Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction","submittedOnDailyBy":{"_id":"6433bc32a4c9c55871a42812","avatarUrl":"/avatars/c27c063d02419b8ba32657514c07e3b2.svg","isPro":false,"fullname":"qian#143","user":"qian43","type":"user","name":"qian43"},"summary":"Streaming 3D reconstruction from extremely long videos requires estimating camera motion and scene geometry online under bounded memory and computation. Early streaming models achieve causal, bounded-cost inference using finite context buffers or compact recurrent states, yet their estimates often deteriorate as sequences grow. Recent methods improve long-horizon stability by coupling short-range context with persistent or multi-level long-range memory. We pursue a different route: we keep the learned temporal state strictly local and formulate predictions whose targets remain independent of sequence length. We present ABot-Recon, a simple streaming model that caches KV features from only the preceding 11 frames. It predicts a point map in the current camera coordinate system together with an adjacent-frame relative pose. These predictions remain equivariant under changes of reference frame, and global poses and geometry are recovered through sequential composition. To reduce accumulated drift, a lightweight temporal refiner improves relative rotations using recent visual and motion context, while a composition-aware pose loss supervises multi-step pose composition. Extensive evaluations on challenging long-sequence benchmarks demonstrate the superior long-horizon performance of our local-context approach. On Oxford Spires, ABot-Recon achieves an ATE of 4.35 m and an RPE-R of 0.12^circ, reducing both errors by approximately 40\\% relative to the best prior results.","upvotes":26,"discussionId":"6a94e198073195fee5157206","projectPage":"https://amap-cvlab.github.io/ABot-Recon-html/","githubRepo":"https://github.com/amap-cvlab/ABot-Recon","githubRepoAddedBy":"user","ai_summary":"ABot-Recon achieves stable long-horizon streaming 3D reconstruction by using only local temporal context and frame-independent predictions composed sequentially, reducing drift via a lightweight temporal refiner and composition-aware pose loss.","ai_keywords":["streaming 3D reconstruction","KV features","point map","relative pose","equivariant predictions","sequential composition","temporal refiner","composition-aware pose loss","ATE","RPE-R"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":229,"organization":{"_id":"641415d08900ef6afa2fcb73","name":"acvlab","fullname":"Alibaba AMAP CV Lab","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6414106ce7d5f817d204e160/dfveRtrRy8Xn7QpG684zl.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6433bc32a4c9c55871a42812","avatarUrl":"/avatars/c27c063d02419b8ba32657514c07e3b2.svg","isPro":false,"fullname":"qian#143","user":"qian43","type":"user"},{"_id":"6566fcb5118497d0af91dc3b","avatarUrl":"/avatars/e081d7d1b1657245cd818e5417cdcb2e.svg","isPro":false,"fullname":"TIANJIAN OUYANG","user":"Ouyangtj","type":"user"},{"_id":"65268d69aa6c7ddb022dc62f","avatarUrl":"/avatars/4202995c3988900b61dc67e0266ffa0a.svg","isPro":false,"fullname":"Wu","user":"haoshu","type":"user"},{"_id":"64a7eaa9ef22f9c793e055a4","avatarUrl":"/avatars/91d71d9ecc49b4caad99a5dbaf3563e7.svg","isPro":false,"fullname":"tian","user":"tianjy05","type":"user"},{"_id":"6645bee6348872b6932db0a6","avatarUrl":"/avatars/f5e8709d57bf944d312f34d2e06673a9.svg","isPro":false,"fullname":"slz","user":"lincent","type":"user"},{"_id":"6620b983c7f53e39e530caab","avatarUrl":"/avatars/4b4891993c0ec3ee76c43ef63a727e38.svg","isPro":false,"fullname":"Chen Yang","user":"cyang203912","type":"user"},{"_id":"62307f6f2c845f238e3ebc6f","avatarUrl":"/avatars/6b5e940c32036fa26eb71730cea8b1bb.svg","isPro":false,"fullname":"XU HU","user":"HXX","type":"user"},{"_id":"69d74ade3af4a5de2c74b1a1","avatarUrl":"/avatars/60c8678d4d24f55d4098b45957c61cbf.svg","isPro":false,"fullname":"chengwei","user":"chengwei-amap","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"6a2da6c8ca070ee12c6e396c","avatarUrl":"/avatars/0355287dcabaa67dbc7f0b10b87451f9.svg","isPro":false,"fullname":"Joe Mama","user":"JoeMama123123123","type":"user"},{"_id":"64b22c65372d4340774373b5","avatarUrl":"/avatars/b9bd0d4e34e4eeb229c0d01aaf601c00.svg","isPro":false,"fullname":"Jincheng Xiong","user":"venite","type":"user"},{"_id":"659bb7700c972f4c775cef97","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/teq47yK1gl_Eu4jii0Ow2.jpeg","isPro":false,"fullname":"Zhiyang Zhang","user":"zzzzzzy123","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":3,"organization":{"_id":"641415d08900ef6afa2fcb73","name":"acvlab","fullname":"Alibaba AMAP CV Lab","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6414106ce7d5f817d204e160/dfveRtrRy8Xn7QpG684zl.png"},"query":{}}">

Papers

arxiv:2608.27529

Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction

Published on Aug 27

· Submitted by

qian#143 on Aug 31

#3 Paper of the day

Alibaba AMAP CV Lab

Upvote

Authors:

Abstract

ABot-Recon achieves stable long-horizon streaming 3D reconstruction by using only local temporal context and frame-independent predictions composed sequentially, reducing drift via a lightweight temporal refiner and composition-aware pose loss.

Generated by thinkingmachines/Inkling-Small

Streaming 3D reconstruction from extremely long videos requires estimating camera motion and scene geometry online under bounded memory and computation. Early streaming models achieve causal, bounded-cost inference using finite context buffers or compact recurrent states, yet their estimates often deteriorate as sequences grow. Recent methods improve long-horizon stability by coupling short-range context with persistent or multi-level long-range memory. We pursue a different route: we keep the learned temporal state strictly local and formulate predictions whose targets remain independent of sequence length. We present ABot-Recon, a simple streaming model that caches KV features from only the preceding 11 frames. It predicts a point map in the current camera coordinate system together with an adjacent-frame relative pose. These predictions remain equivariant under changes of reference frame, and global poses and geometry are recovered through sequential composition. To reduce accumulated drift, a lightweight temporal refiner improves relative rotations using recent visual and motion context, while a composition-aware pose loss supervises multi-step pose composition. Extensive evaluations on challenging long-sequence benchmarks demonstrate the superior long-horizon performance of our local-context approach. On Oxford Spires, ABot-Recon achieves an ATE of 4.35 m and an RPE-R of 0.12^circ, reducing both errors by approximately 40\% relative to the best prior results.