Frustum PointNets is a 3D object detection system that combines 2D object detectors with 3D deep learning. The pipeline uses 2D bounding boxes from RGB images to define 3D frustum regions, then applies PointNet/PointNet++ networks on point clouds within those regions to perform 3D instance segmentation and amodal bounding box estimation. It works directly on 3D point clouds without voxelization, leveraging coordinate normalization techniques for improved learning.