EFFOcc: A Minimal Baseline for EFficient Fusion-based 3D Occupancy Network

We provide lidar-camera occupancy prediction video of Occ3D-nuScenes dataset.

EFFOcc explores towards the minimal (minimal computation costs and minimal label costs) baseline for fast and high-performance 3D occupancy prediction with lidar-camera fusion. We show with proper detection pretraining, lightweight BEV-based fusion occnet can perform as well as voxel-based fusion occnets. Then, We conduct activate learning with maximum entropy on frame- and voxel-level to see the minimum label requirements for occupancy prediction.

EFFOcc on Occ3D-nuScenes dataset: EFFOcc on Occ3D-Waymo dataset: EFFOcc on OpenOccupancy-nuScenes dataset:

EFFOcc on two-stage active learning setting:

Thanks to prior excellent open source projects:

• BEVDet
• OpenOccupancy
• FlashOcc
• CRB-active-3Ddet