FEANet: Foreground-edge-aware network with DenseASPOC for human parsing

Human parsing has drawn a lot of attention from the public due to its critical role in high-level computer vision applications. Recent works demonstrated the effectiveness of utilizing context module and additional information in improving the performance of human parsing. However, ambiguous objects, small scaling and occlusion problems are still the bottlenecks. In this paper, we propose a novel framework called Foreground-Edge-Aware Network (FEANet) with DenseASPOC context module to further enhance the segmentation performance for human parsing. We claim that the fusion of foreground and edge information can effectively segment occluded regions by reducing the impact of pixels occupied by non-human object parts while persevering boundaries between each class. Moreover, we introduce the Dense Atrous Spatial Pyramid Object Context (DenseASPOC) module to address the problem of small and ambiguous objects by empowering feature extraction ability with solid spatial perception and semantic context information. We conducted comprehensive experiments on various human parsing benchmarks including both single-human and multi-human parsing. Both quantitative and qualitative results show that the proposed FEANet has superiority over the current methods. Moreover, detailed ablation studies report the effectiveness of the employment on each contribution. (c) 2021 Elsevier B.V. All rights reserved.