Vision
I work toward geometrically grounded world models — 3D representations that machines can not only recover from raw observations, but roll forward in time and condition on action.
In the short term, this means tightening the foundations of 3D reconstruction (point cloud registration, scene representation, robust geometry). In the longer term, it means giving those representations dynamics — so a machine can not only see a scene, but predict how it evolves.
Current Foundations
My published work centers on three building blocks of a usable 3D world:
- Robust geometry & estimation. Real sensors produce noisy, partial, contaminated data. TurboReg (ICCV 2025) and HeMoRa (CVPR 2025) push robust estimators for point cloud registration onto a faster, more hardware-aware footing.
- Structured matching. Correspondences across views are the connective tissue of any 3D pipeline. ML-SemReg (ECCV 2024) shows that adding lightweight semantic structure can sharply lift match recall — without any learning.
- Scene-grounded perception. With ULF-Loc (CVPR 2026 Highlight), I started moving from “geometry under registration” to “geometry under representation” — localizing cameras against 3D Gaussian Splatting scenes via unbiased landmark features.
Where It’s Going
Three threads I want to follow during my PhD:
- From static to dynamic — recovering not just frozen scenes but how they move. Dynamic 3DGS, deformable scene representations, and 4D world models.
- From observation to prediction — modeling temporal evolution and action-conditioned rollouts on top of geometric representations.
- From perception to use — testing world models where they actually matter: robotic manipulation, autonomous driving, embodied agents.
The bet I am making: strong geometric grounding will be a feature, not a constraint, for the next generation of world models — especially in robotics and autonomous driving, where ungrounded generation cannot be trusted.
Selected Papers
For the full list, see publications on the home page. Highlights:
- ULF-Loc — Unbiased Landmark Feature for Robust Visual Localization with 3D Gaussian Splatting — CVPR 2026 Highlight.
- TurboReg — TurboClique for Robust and Efficient Point Cloud Registration — ICCV 2025.
- HeMoRa — Unsupervised Heuristic Consensus Sampling for Robust Point Cloud Registration — CVPR 2025.
- ML-SemReg — Boosting Point Cloud Registration with Multi-level Semantic Consistency — ECCV 2024.
Always open to collaboration — shaochengyan@whu.edu.cn.