(An) large-scale point cloud deep neural network accelerator with virtual pillar and ROI-based skipping

Abstract

This paper proposes a large-scale point cloud deep neural network accelerator. It presents a novel concept of virtual pillars and proposes a skipping method based on the region of interest. Additionally, it maximizes slice-level sparsity by adopting the sign-magnitude representation method instead of the conventional two's complement representation. This work achieves a speedup of 5.3 times on feature encoding and shows 35.4 TOPS/W power efficiency on backbone.