PointPillars: Fast Encoders for Object Detection From Point Clouds

PointPillars tackles how to encode a lidar point cloud into a representation suitable for a downstream object detection pipeline, a step where prior fixed encoders were fast but less accurate and learned encoders were accurate but slow. The method uses PointNets to learn a representation of points organized into vertical columns called pillars; the resulting encoded features can be consumed by any standard 2D convolutional detection architecture, and the authors add a lean downstream network.

Extensive experiments show that PointPillars outperforms previous encoders on both speed and accuracy by a large margin, and the full detection pipeline, using only lidar, significantly exceeds the state of the art, even compared with fusion methods, on the 3D and bird's-eye-view KITTI benchmarks. It achieves this while running at 62 Hz, a two-to-four-fold runtime improvement, with a faster variant matching prior state of the art at 105 Hz, making it a practical encoding for robotics and autonomous driving.