| Input Type |
3D voxels from LiDAR point cloud |
Point clouds projected onto pillars |
BEV projected images or LiDAR points |
Multi-modal: LiDAR + camera + radar |
| Representation |
3D voxel grid |
2D pseudo-image (pillar features) |
2D BEV feature map |
2D BEV feature map with fused modalities |
| Computational Cost |
High (3D convolutions) |
Moderate (2D convolutions) |
Low to moderate (2D CNN) |
High (fusion + multi-modal processing) |
| Spatial Resolution |
Good 3D spatial accuracy |
Limited by pillar size |
Good horizontal resolution, limited vertical |
High overall, retains vertical + semantic info |
| Detection Performance |
Strong for dense point clouds |
Efficient and competitive |
Efficient, good for overhead view |
Best accuracy, robust to occlusion |
| Use Cases |
3D object detection in LiDAR |
Real-time LiDAR detection |
Autonomous driving, traffic analysis |
Autonomous driving, robust multi-sensor perception |
| Pros |
Accurate 3D localization |
Fast, lightweight |
Easy to integrate with planning |
Robust, handles occlusion and multi-modal info |
| Cons |
Computationally heavy |
Loss of fine 3D detail |
Limited vertical info |
More complex, high compute requirements |
