PointPillars: Fast Encoders for Object Detection From Point Clouds
Introduces PointPillars, a point cloud encoder using PointNets over vertical pillars for fast, accurate 3D object detection from lidar.
PointPillars is an encoder that converts a point cloud into a form suited to downstream 3D object detection, addressing the trade-off between fast fixed encoders and accurate but slow learned ones. It uses PointNets to learn features from points organized into vertical columns, or pillars, whose output feeds a standard 2D convolutional detection network. Experiments show PointPillars surpasses prior encoders in both speed and accuracy, and its lidar-only pipeline beats the state of the art on the KITTI 3D and bird's-eye-view benchmarks while running at 62 Hz.
Based on: PointPillars: Fast Encoders for Object Detection From Point Clouds · Computer Vision and Pattern Recognition
Curated by Aramai Editorial
Read summary →