BiEPNet: Bilateral Edge-perceiving Network for High-Resolution Human Parsing

Human parsing is a fundamental task aimed at segmenting human images into distinct body parts and holds vast potential applications. Nowadays, the advancement of image-capturing devices has led to a growing number of high-resolution human images. Receptive field, detail loss and memory usage are a triplet of contradictions in high-resolution scenarios. Existing human parsing methods designed for low-resolution inputs struggle to process high-resolution images efficiently due to their massive demands for computation and memory. Some methods save resources by overwhelmingly downsampling or encoding high-resolution inputs at the cost of poor performance on details. To resolve the issues above, we propose the Bilateral Edge-Perceiving Network (BiEPNet), consisting of a resources-friendly semantic-perceiving branch to acquire sufficient global information and a simple yet effective edge-perceiving branch used to refine details. The attention mechanism is utilized to simultaneously enhance the perception of context and details, leading to better performance on the boundary regions. To verify the effectiveness of BiEPNet, we contribute a high-resolution human parsing dataset, Human4K, containing 4,000 images with more than five million pixels. Extensive experiments on Human4K demonstrate that our method effectively outperforms the state-of-the-art methods.