(A) low-power graph convolutional network processor for real-time 3D point cloud semantic segmentation

Abstract

A low-power graph convolutional network (GCN) is proposed for accelerating 3D point cloud semantic segmentation (PCSS) in real-time on mobile devices. Three key features enable the low power GCN-based 3D PCSS. First, the new hardware friendly GCN algorithm, sparse grouping based dilated graph convolution (SG-DGC) is proposed and reduce power consumption through the reduction of computation and external memory access (EMA). SG-DGC reduces 71.7% of the overall computation and 76.9% of EMA through the sparse grouping of the point cloud. Proposed processor accelerate the GCN with the two-level pipeline (TLP) consisting of the point-level pipeline (PLP) and group-level pipelining (GLP). The PLP enables point-level module-wise fusion (PMF) which reduces 47.4% of EMA with minimal additional on-chip memory footprint through point-level execution. The GLP increased the core utilization by 21.1% through balancing the workload of graph generation and graph convolution with the loosened data dependency through SG-DGC and enable 1.1× higher throughput. Finally, center point feature reuse (CPFR) reuse computation results of the redundant operation and reduces 11.4% of computation. The processor is implemented with 65nm CMOS technology, and the 4.0mm$^2$ 3D PCSS processor show 95mW power consumption while operating in real-time of 30.8 frames-per-second 3D PCSS with 4k points.