VGOcc Learns Visual-Geometric Gaussians for Vision-Centric 3D Driving Occupancy PredictionVGOcc：面向纯视觉自动驾驶3D占用预测的视觉-几何高斯学习方法
S 1.7T11 sources1 个来源R7-research
Researchers Junhong Lin, Xianda Guo, Kangli Wang, Yuqi Ye, Xiaoyu Liang, Yanlun Peng and Wei Gao posted VGOcc (arXiv:2607.18078v1) on 20 Jul 2026, targeting vision-only 3D occupancy prediction that recovers a semantic 3D occupancy field from calibrated surround-view images despite depth ambiguity along each camera ray.
The paper notes prior methods progressed from dense structured representations to sparse Gaussian primitives to improve 3D scene-representation efficiency, but states Gaussian learning still relies mainly on image-domain features that offer limited explicit geometric information for volumetric reconstruction — the gap VGOcc's visual-geometric Gaussian learning aims to close.
Vision-only occupancy prediction from surround-view cameras is central to camera-based autonomous-driving perception stacks, letting vehicles infer 3D free space and obstacles without relying on lidar or other explicit depth sensors.