NeurIPS 2026 · First author. CitySTAR grounds natural-language queries in city-scale 3D scenes through structured constraints, scene graphs and topology-aware reasoning.
ECCV 2026 · Co-author. SkyLume benchmarks urban 3D reconstruction across changing illumination, pairing multi-temporal aerial imagery with LiDAR ground truth.
Geometry-guided RGB-D projection and visual foundation models enable training-free open-vocabulary 3D point-cloud segmentation.
CVPR 2026 Findings · First author. UniD-Shift separates shared and modality-specific features for interpretable 2D–3D semantic segmentation and cross-domain generalization.
This paper proposes SPDC, a dual-scale contrastive learning network combining super-point and point-level features for self-supervised point cloud semantic segmentation, achieving …