Our paper on vision-based 3D occupancy prediction is accepted to IEEE RA-L
[2026.10.14]
The following paper is accepted to the IEEE Robotics and Automation Letters (RA-L):
GAP3D: Geometry-Aware Adaptive Planar Representations for 3D Occupancy Prediction by Hongjung Lee, Jae Goo Choy, Howoong Jun, Jeongwoo Oh, and Songhwai Oh
- Abstract: Reliable 3D occupancy prediction is essential for mobile robots and autonomous vehicles, where navigation and collision avoidance require reasoning about occupied and free space in the surrounding environment. Existing methods typically use dense voxel grids, compact feature maps, or sparse 3D representations, but driving scenes exhibit strongly non-uniform geometry along the vertical axis. We propose Geometry-Aware Adaptive Planar Representations (GAP3D), a structured representation that preserves regular horizontal coverage while adapting its vertical feature support to the observed scene. GAP3D constructs a small set of adaptive planes whose heights vary with the observed scene geometry, allowing vertical feature support to be allocated where it is most informative. Geometry-aware splatting then maps the resulting plane features into a fixed dense occupancy volume using Gaussian weights determined by their learned heights and spreads. This enables the intermediate representation to redistribute its limited vertical support according to scene geometry while retaining regular horizontal structure. GAP3D achieves 42.8 RayIoU on Occ3D-nuScenes at 9.0 FPS on an NVIDIA A100, outperforming the compared methods, and obtains strong results on SurroundOcc and OpenOccupancy. Controlled ablations further isolate the contributions of adaptive plane placement and geometry-aware splatting.