SLICENet: An FPGA-Based Efficient Semantic Segmentation Network for Edge Deployment

Abstract

Semantic segmentation is a pixel-level visual recognition task widely used in autonomous driving. Attaining a decent trade-off between accuracy and speed is critically important for the effective physical deployment of networks on resource-constrained edge devices. Towards this challenging task, we propose an efficient basic block that is designed to leverage local, short-range, and long-range contextual information at different abstraction levels. We introduce a simple technique inside the basic block, called Iterative Context Embedding (ICE), to reinforce the short and long-range contextual details in an iterative fashion. Based on the resulting short and long-range ICE or SLICE module, we propose an ultra-lightweight network, called SLICENet. Our model is the fastest among the existing ultra-lightweight models while achieving a decent accuracy. Specifically, with only 0.3 million parameters, it achieves 69.1% mean IoUs on the cityscapes test set, making it the smallest model to achieve this accuracy. In addition, it achieves an inference speed of 224.8 frames per second (FPS) on the RTX 3090 with $512\times 1024$ resolution. To achieve a power-efficient solution meant for battery-operated devices, we also deploy our model on Xilinx’s ZCU102 development board (Zync UltraScale+ MPSoC). Despite achieving an impressive performance, its power consumption is only 950 mW; significantly lower than GPU-based inferences. Our code will be shared at https://github.com/NadeemAtif-Alig/SLICENet.