WITHDRAWN

Abstract:

Object detection in LiDAR point clouds is an important application especially for autonomous driving. So far, most work has been focused to develop deep learning architectures and to improve the detection performance. However, in addition to accuracy, inference speed is a very important component to quickly detect moving objects.  We present an evaluation and comparison of the inference of three state-of-the-art deep learning models for LiDAR-based 3D object detection: PointPillars, PV-RCNN and CenterPoint. These models were created using the OpenPCDet library. We describe the steps necessary to convert the PyTorch models to Torch-Script and how to optimize them for the TensorRT runtime. Using mix precision with PyTorch models and half-precision Torch-Script models significantly reduces the runtime while having very little impact on the detection performance. All the experiments were performed on the KITTI benchmark dataset and show that the proposed approach achieves significant speed-ups.