Excited to share our new work: Self-Geometry! 🎉</p>\n<p>TL;DR: A GT-free, plug-and-play test-time adaptation pipeline that imposes explicit multi-view geometric constraints on pretrained 3D Vision Foundation Models, using 2D pixel correspondences as pseudo ground-truth.</p>\n<p>Prior TTA methods rely on implicit self-consistency derived from the model's own outputs, which barely helps exactly where the pretrained VFM is most inaccurate. We supervise with explicit geometry instead, and adapt only LoRA parameters, completing per-scene adaptation in a few minutes on a single GPU.</p>\n<p>Result: consistent mean improvements in both pose and geometry estimation across 6 VFMs (VGGT, π³, DA3-Giant/Large/Base/Small) and 4 benchmarks (7Scenes, ETH3D, ScanNet++, HiRoom).</p>\n<p>📄 arXiv: <a href=\"https://arxiv.org/abs/2608.10708\" rel=\"nofollow\">https://arxiv.org/abs/2608.10708</a><br>🌐 Project page (with interactive side-by-side comparisons): <a href=\"https://cmlab-korea.github.io/Self-Geometry/\" rel=\"nofollow\">https://cmlab-korea.github.io/Self-Geometry/</a></p>\n<p>Happy to answer any questions. Feedback and discussion very welcome!</p>\n","updatedAt":"2026-08-13T03:30:43.481Z","author":{"_id":"693a2cd463a8ded7071c6365","avatarUrl":"/avatars/3befa67c2bf7e9c3cb376f6b6f227242.svg","fullname":"Seokhyun Youn","name":"Seokhyun01","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8768990635871887},"editors":["Seokhyun01"],"editorAvatarUrls":["/avatars/3befa67c2bf7e9c3cb376f6b6f227242.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.10708","authors":[{"_id":"6a7d36760ac8bee77474eeb0","user":{"_id":"693a2cd463a8ded7071c6365","avatarUrl":"/avatars/3befa67c2bf7e9c3cb376f6b6f227242.svg","isPro":false,"fullname":"Seokhyun Youn","user":"Seokhyun01","type":"user","name":"Seokhyun01"},"name":"Seokhyun Youn","status":"claimed_verified","statusLastChangedAt":"2026-08-13T08:45:05.033Z","hidden":false},{"_id":"6a7d36760ac8bee77474eeb1","name":"Dahyeon Kye","hidden":false},{"_id":"6a7d36760ac8bee77474eeb2","name":"Sung-Ho Bae","hidden":false},{"_id":"6a7d36760ac8bee77474eeb3","name":"Jihyong Oh","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/693a2cd463a8ded7071c6365/weWtJWWHarpzDKV-SR0UD.mp4"],"publishedAt":"2026-08-11T00:00:00.000Z","submittedOnDailyAt":"2026-08-13T00:00:00.000Z","title":"Self-Geometry: GT-Free and Plug-and-Play Test-Time Adaptation for Geometrically Consistent 3D Vision Foundation Models","submittedOnDailyBy":{"_id":"693a2cd463a8ded7071c6365","avatarUrl":"/avatars/3befa67c2bf7e9c3cb376f6b6f227242.svg","isPro":false,"fullname":"Seokhyun Youn","user":"Seokhyun01","type":"user","name":"Seokhyun01"},"summary":"Recent Vision Foundation Models (VFMs) predict depth, camera pose, and pointmap in a single forward pass without per-scene optimization, achieving strong generalization. However, enforcing explicit multi-view geometric consistency, e.g., through bundle adjustment, is computationally costly and is thus not imposed during VFM pretraining, so such inconsistency can arise. To address this, implicit self-consistency derived from model outputs (e.g., pointmaps, features), though enforced at test-time in prior work, delivers inherently limited performance gain, especially on scenes where the pretrained VFM is highly inaccurate. In contrast to this implicit signal, we propose Self-Geometry, a plug-and-play test-time adaptation pipeline that directly imposes explicit multi-view geometric constraints using 2D pixel correspondences as pseudo ground-truth. Our proposed Self-Geometry consists of Geometric Disentanglement Optimization, which combines Multi-View Consistency and Epipolar Consistency losses with Gradient Disentanglement to prevent gradient conflict; Frame Angular-Neighbor, a view sampler based on SO(3) geodesic distances for lightly imposing these constraints; and Lightweight TTA, which adapts VFMs via LoRA. Our method achieves consistent improvements in both pose and geometry estimation across six VFMs (VGGT, π^3, DA3-Giant/Large/Base/Small) and four benchmarks (7Scenes, ETH3D, ScanNet++, HiRoom).","upvotes":11,"discussionId":"6a7d36760ac8bee77474eeb4","projectPage":"https://cmlab-korea.github.io/Self-Geometry/","githubRepo":"https://github.com/CMLab-Korea/Self-Geometry","githubRepoAddedBy":"user","ai_summary":"Self-Geometry improves vision foundation model predictions by enforcing explicit multi-view geometric constraints via test-time adaptation with LoRA, disentangled losses, and angular neighbor sampling.","ai_keywords":["Vision Foundation Models","multi-view geometric consistency","bundle adjustment","Self-Geometry","Geometric Disentanglement Optimization","Multi-View Consistency loss","Epipolar Consistency loss","Gradient Disentanglement","Frame Angular-Neighbor","SO(3) geodesic distances","Lightweight TTA","LoRA","pointmaps"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":1,"organization":{"_id":"668d03f09761c585a203385f","name":"Chung-AngUniversity","fullname":"Chung-Ang University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/668d037bb09a05f3d78d5b4c/eMdFRsUfQGWwCoPeQ3mkH.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6937b46ec84bb51be05c6908","avatarUrl":"/avatars/85a853b4cc382b443020db48144ad1fa.svg","isPro":false,"fullname":"Jeahun Sung","user":"jeahun","type":"user"},{"_id":"6530d878f145530101ef3e4f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6530d878f145530101ef3e4f/T9ldn9YSQcTXMjdAr7B7X.jpeg","isPro":false,"fullname":"Choi Seungho","user":"ohgnues","type":"user"},{"_id":"69aeafefdde330cbe80c5e01","avatarUrl":"/avatars/4f9e5b92a242406436c341801f5268b8.svg","isPro":false,"fullname":"S","user":"Dovakho014","type":"user"},{"_id":"66820b2d2455dc29be68663d","avatarUrl":"/avatars/610661b7dd6d887be36a15ed5425bad3.svg","isPro":false,"fullname":"Dahyeon Kye","user":"kyeda","type":"user"},{"_id":"6039478ab3ecf716b1a5fd4d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6039478ab3ecf716b1a5fd4d/_Thy4E7taiSYBLKxEKJbT.jpeg","isPro":true,"fullname":"taesiri","user":"taesiri","type":"user"},{"_id":"660d784051ed58ee308b6ac9","avatarUrl":"/avatars/a9b252e947a2a22c81f2302c4401e1b4.svg","isPro":false,"fullname":"Sukhun","user":"Kukohun","type":"user"},{"_id":"634cfebc350bcee9bed20a4d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/634cfebc350bcee9bed20a4d/fN47nN5rhw-HJaFLBZWQy.png","isPro":false,"fullname":"Xingyi Yang","user":"adamdad","type":"user"},{"_id":"631c386bc73939ffc0716a37","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1662793811119-noauth.jpeg","isPro":false,"fullname":"SeongWan Kim","user":"idgmatrix","type":"user"},{"_id":"65c4eb7cd1dcbd30d86febec","avatarUrl":"/avatars/001c8f02e8ce794b2c21883628b2da72.svg","isPro":false,"fullname":"free-bit","user":"free-bit","type":"user"},{"_id":"664d930f4b870dd167473c1c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/664d930f4b870dd167473c1c/TXVEPGvkhftdI_xE1mluu.jpeg","isPro":false,"fullname":"Andy Guan","user":"andytonglove","type":"user"},{"_id":"64e9f5bbf494f8b2a0670eee","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/rM0q9i3DXc7nVY6MDjiCY.jpeg","isPro":false,"fullname":"Rijusmit Biswas","user":"Phantomcloak19","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"668d03f09761c585a203385f","name":"Chung-AngUniversity","fullname":"Chung-Ang University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/668d037bb09a05f3d78d5b4c/eMdFRsUfQGWwCoPeQ3mkH.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.10708.md","query":{}}">
Self-Geometry: GT-Free and Plug-and-Play Test-Time Adaptation for Geometrically Consistent 3D Vision Foundation Models
Abstract
Self-Geometry improves vision foundation model predictions by enforcing explicit multi-view geometric constraints via test-time adaptation with LoRA, disentangled losses, and angular neighbor sampling.
Recent Vision Foundation Models (VFMs) predict depth, camera pose, and pointmap in a single forward pass without per-scene optimization, achieving strong generalization. However, enforcing explicit multi-view geometric consistency, e.g., through bundle adjustment, is computationally costly and is thus not imposed during VFM pretraining, so such inconsistency can arise. To address this, implicit self-consistency derived from model outputs (e.g., pointmaps, features), though enforced at test-time in prior work, delivers inherently limited performance gain, especially on scenes where the pretrained VFM is highly inaccurate. In contrast to this implicit signal, we propose Self-Geometry, a plug-and-play test-time adaptation pipeline that directly imposes explicit multi-view geometric constraints using 2D pixel correspondences as pseudo ground-truth. Our proposed Self-Geometry consists of Geometric Disentanglement Optimization, which combines Multi-View Consistency and Epipolar Consistency losses with Gradient Disentanglement to prevent gradient conflict; Frame Angular-Neighbor, a view sampler based on SO(3) geodesic distances for lightly imposing these constraints; and Lightweight TTA, which adapts VFMs via LoRA. Our method achieves consistent improvements in both pose and geometry estimation across six VFMs (VGGT, π^3, DA3-Giant/Large/Base/Small) and four benchmarks (7Scenes, ETH3D, ScanNet++, HiRoom).
Community
Excited to share our new work: Self-Geometry! 🎉
TL;DR: A GT-free, plug-and-play test-time adaptation pipeline that imposes explicit multi-view geometric constraints on pretrained 3D Vision Foundation Models, using 2D pixel correspondences as pseudo ground-truth.
Prior TTA methods rely on implicit self-consistency derived from the model's own outputs, which barely helps exactly where the pretrained VFM is most inaccurate. We supervise with explicit geometry instead, and adapt only LoRA parameters, completing per-scene adaptation in a few minutes on a single GPU.
Result: consistent mean improvements in both pose and geometry estimation across 6 VFMs (VGGT, π³, DA3-Giant/Large/Base/Small) and 4 benchmarks (7Scenes, ETH3D, ScanNet++, HiRoom).
📄 arXiv: https://arxiv.org/abs/2608.10708
🌐 Project page (with interactive side-by-side comparisons): https://cmlab-korea.github.io/Self-Geometry/
Happy to answer any questions. Feedback and discussion very welcome!
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.10708 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.10708 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.10708 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.