A Review of PointPillars Architecture for Object Detection from Point Clouds

Object detection from point clouds, e.g. 3D LiDAR data, has many applications especially for autonomous driving systems. There have been several approaches to this complex problem, however, PointPillars architecture has the advantage that it can reuse the existing image-based convolution neural networks for object detection from 3D LiDAR data. The performance of the PointPillars architecture is further enhanced in this work by introduction of an additional Pillar Feature Extraction layer. It is observed that this modified PointPillars model trained to detect cars on KITTI dataset shows an improvement of 6.25% in the average precision for the easy cases in KITTI 3D detection benchmark, when tested on a GTX 1080i GPU.