arxivcs.CV2026-07-20
VGOcc: Learning Visual-Geometric Gaussians for Vision-Centric 3D Driving Occupancy Prediction
Junhong Lin, Xianda Guo, Kangli Wang, Yuqi Ye, Xiaoyu Liang, Yanlun Peng, et al.
Vision-only occupancy prediction requires recovering a semantic 3D occupancy field from calibrated surround-view images, where each view provides observations with ambiguous depth along camera rays. Existing methods have progressed from dense structured representations to sparse…