Pruning CNNs for LiDAR-based Perception in Resource Constrained Environments

Deep neural networks provide high accuracy for perception. However they require high computational power. In particular, LiDAR-based object detection delivers good accuracy and real-time performance, but demands high computation due to expensive feature-extraction from point cloud data in the encoder and backbone networks. We investigate the model complexity versus accuracy trade-off using reinforcement learning based pruning for PointPillars, a recent LiDAR-based 3D object detection network. We evaluate the model on the validation dataset of KITTI (80/20-splits) according to the mean average precision (mAP) for the car class. We prune the original PointPillars model (mAP 89.84) and achieve 65.8% reduction in floating point operations (FLOPs) for a marginal accuracy loss. The compression corresponds to 31.7% reduction in inference time and 35% reduction in GPU memory on GTX 1080 Ti.