TinyPillarNet: Tiny Pillar-Based Network for 3D Point Cloud Object Detection at Edge

Limited by huge computational cost, high inference latency and large memory consumption, existing 3D point cloud object detection methods are hard to be deployed on Internet of Things (IoT) edge devices. To handle this challenge, we present an extremely tiny framework termed TinyPillarNet. This framework leverages innovative pillar encoder to represent point cloud as immensely tiny pseudo-maps for extremely shrinking the input 3D sensing data. Moreover, a compact dual-stream feature extraction network is put forward to respectively extract intrinsic feature and distributional saliency map, which jointly boosts the detection precision with the lowest hardware cost. Extended experiments on KITTI benchmark demonstrated that our TinyPillarNet yields applicable precision with a record tiny weight size of 1.69 MB at a high inference speed of 1.67 times faster than the current record. Furthermore, the specially designed prototype verification system achieves a superior energy efficiency, which outperforms the similar deep learning based point cloud processing solutions on FPGA with a big margin.