Low-latency Human Portrait Segmentation Network Optimized for CPU Inference

Abstract

In this paper, we discuss a design of fast and lightweight neural networks for working in real-time under very strict resource constraints and describe a human portrait segmentation method with temporal consistency based on an encoder-decoder architecture with a state-of-the-art CPU optimized PP-LCNet backbone and a custom decoder. Proposed neural network can process about 150-500 frames per second using only a single CPU thread with high accuracy and can be used for virtual background replacements in video conferencing and other augmented reality cases