Tuan Duc Ngo

Reconstructs the complete 3D scene from a single RGB image, including surfaces hidden behind visible geometry, by generating a structured volumetric representation instead of regressing per-pixel depth.

Recovers sharp, view-consistent geometry and accurate camera poses from uncalibrated videos at up to 2K resolution, using a dual-stream transformer that separates global context from fine detail.

Tracks every pixel in 3D across long monocular videos, running over 8× faster than prior methods while achieving state-of-the-art dense 2D and 3D tracking accuracy.

Segments 3D object instances from open-vocabulary queries by lifting 2D instance masks across frames into 3D proposals, recovering small and geometrically ambiguous objects that 3D-only methods miss.

Segments 3D point cloud instances without clustering, using instance-aware sampling and box-aware dynamic convolution, and sets new state-of-the-art results on ScanNetV2, S3DIS, and STPLS3D.