Multi-Scale PointPillars 3D Object Detection Network

Point cloud object detection is a key step in many 3D applications, such as autonomous driving and housekeeping robots. In this paper, we present MPNet, a Multi-Scale PointPillars Network for 3D object detection using point clouds. This paper aims at high accuracy and efficient speed. The proposed neural network is mainly composed of three part, a feature extracting part, a region proposal network, and finally a detection head. We use multi-scale encoders to concurrently extract features by PointNet, and then treat the 3D feature map as a pseudo image with many channels. Inspired by feature pyramid networks (FPN) and deformable convolution, our region proposal network achieves high accuracy while keeps high speed. MPNet combines features with different receptive field, and accelerates model by using 2D convolution, balancing the accuracy and speed well. Experiments on KITTI car detection benchmark show that our MPNet captures better location information on the bird's eye view. It ranks the third place in the Bird's Eye View Leaderboard. Meanwhile, it is the most efficient model compared within the top 10 models in the Bird's Eye View leaderboard.