frustum-pointnets

Why we included this project

This is the reference implementation of Frustum PointNets, the CVPR 2018 approach to detecting objects in 3D point clouds, released by the Stanford and Nuro team behind the paper. The pipeline runs a mature 2D detector on RGB images, uses each bounding box to carve out a 3D frustum, then applies PointNet-style segmentation and amodal bounding box estimation directly to the points inside. Because it operates on raw point clouds instead of voxel grids, it keeps the full 3D geometry while leaning on image cues to shrink the search space. Evaluated on KITTI and SUNRGBD, it still holds up as a solid reference point for teams building LiDAR or depth-based detection for autonomous driving, robotics, and 3D scene understanding.