RayOcc Reframes Camera-Only 3D Occupancy Prediction as Multi-Label Ray Modeling to Tackle OcclusionRayOcc：将摄像头3D占用预测重构为多标签射线建模，破解遮挡难题
S 1.7T11 sources1 个来源R7-research
Junho Kim and Seongwon Lee (arXiv:2607.17660v1, submitted July 20, 2026) propose RayOcc, a camera-only 3D semantic occupancy prediction method that reformulates the task from single-depth estimation into a multi-label existence problem along each camera ray.
The method uses Gaussian Mixture Intensity to model multiple spatially separated surfaces coexisting along a single ray, addressing the depth ambiguity and occlusion that limit prior approaches, which favor only one dominant depth hypothesis per ray.
By explicitly modeling occluded, multi-surface geometry rather than a single depth per ray, RayOcc targets a core bottleneck in camera-based autonomous driving perception, where accurately inferring hidden or overlapping obstacles is essential for downstream path planning.
Junho Kim和Seongwon Lee（arXiv:2607.17660v1，2026年7月20日提交）提出RayOcc，将摄像头3D语义占用预测从单深度估计重新表述为沿每条射线的多标签存在性问题。