[6/8] Motion-Refined DINOSAUR for Unsupervised Multi-Object Discovery (Oral at ILR+G Workshop)

by Xinrui Gong*, @olvrhhn.bsky.social *, @christophreich.bsky.social , Krishnakant Singh, @simoneschaub.bsky.social , @dcremers.bsky.social @stefanroth.bsky.social

[3/8] Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion

by @jev-aleks.bsky.social *, @christophreich.bsky.social *, @fwimbauer.bsky.social , @olvrhhn.bsky.social , Christian Rupprecht, @stefanroth.bsky.social, @dcremers.bsky.social

🌍 visinf.github.io/scenedino/

🌍Project Page: visinf.github.io/scenedino/
📃Paper: arxiv.org/abs/2507.06230
💻Code: github.com/tum-vision/s...
🤗Demo: huggingface.co/spaces/jev-a...

Work by: @jev-aleks.bsky.social*, @christophreich.bsky.social*, @olvrhhn.bsky.social, Christian Rupprecht, @stefanroth.bsky.social, and @dcremers.bsky.social @ TUM CVG, @visinf.bsky.social, @munichcenterml.bsky.social, @zuseschooleliza.bsky.social, and @hessianai.bsky.social

✅ SceneDINO offers refined, high-resolution, and multi-view consistent (rendered) 2D features.

✅SceneDINO outperforms our unsupervised baseline (S4C + STEGO) in unsupervised SSC accuracy.
✅Linear probing our feature field leads to an SSC accuracy on par with 2D supervised S4C.

⚗️Distilling and clustering SceneDINO's feature field in 3D results in unsupervised semantic scene completion predictions.

🏋SceneDINO is trained to estimate an expressive 3D feature field using multi-view self-supervision and 2D DINO features.

🚀 SceneDINO is unsupervised and infers 3D geometry and features from a single image in a feed-forward manner. Distilling and clustering SceneDINO's 3D feature field lead to unsupervised semantic scene completion predictions.

Scene-Centric Unsupervised Panoptic Segmentation

by @olvrhhn.bsky.social , @christophreich.bsky.social , @neekans.bsky.social , @dcremers.bsky.social, Christian Rupprecht, and @stefanroth.bsky.social

Sunday, 8:30 AM, ExHall D, Poster 330
Project Page: visinf.github.io/cups