3d object detection from point cloud

Abstract

Here we present a comparison of two different deep learning architectures’ effectiveness along with two distinct detection head approaches for detecting point cloud ball objects. Two backbones that are explored are: VoxelNet, suited for organized point clouds, and PointNet, which handles unorganized point clouds. We modified and implemented SSD and Faster R-CNN detection heads for both backbones. It turns out that the PointNet backbone integrated with a customized Faster R-CNN detection head achieved higher accuracy compared to other combination of models.