For human parsing in wild and cluttered environment, deep architectures are widely utilized since they yield strong segmentation performance, while with the price of huge model size and computational complexity. These properties are highly limiting for them if to be deployed on resource-limited platforms, in particular for edge intelligence in real-time. In this study, we propose a lightweight framework for human parsing, named Fast DSPP+PGN+Attn, which focuses on the efficiency-accuracy trade-off. The proposed model consists of a MobileNetV2 back-bone (i.e., the AirLab-Net), a Dilated Spatial Pyramid Pooling (DSPP) block to capture multi-scale contextual information, a pixel grouping decoder employing the PGN for improved part boundary flow consistency and potency spatial and squeeze-and-excitation attention modules are used for feature refinement. However, the model with a relatively compact size—i.e., 2.14M parameters and 5.70 GFLOPs— could achieve the performance of 40.67% mean IoU (mIoU) and 87.3% pixel accuracy on the CIHP benchmark, while running at a speed of 51.9 frames per second on a single GPU. These findings indicate that combining the contextual aggregation methods with the method of structured pixel grouping is a strategy to exploit orthogonal avenues and cross-examine their complementarity, more efficiently and potentially achieve better segmentation quality without lost of real-time performance. Therefore, the proposed method can be widely applicable in embedded vision system, surveillance and mobile perception.
Ouza et al. (Thu,) studied this question.