ClothFormer - A Boundary Aware Self-Attention Network for Human Outfit Parsing

Advanced deep learning-based networks have redefined state-of-the-art performance on parsing tasks such as human parsing, face parsing etc., owing to their potential to extract deep features from significant datasets. One such task in the succession of parsing tasks is human outfit or cloth parsing. The semantic disparity between identical features led to inefficient parsing even after adopting heavy and dense architectures involving deep convolutional neural networks (DCNN), notably at border levels. Feature extraction with variations in pose and illumination had also become some of the challenges of cloth parsing. To address these problems, ClothFormer - An Boundary Aware Self-Attention Network for Human Outfit Parsing - is introduced which explores the use of multi-scale techniques to enhance the per-pixel frames' reliability and accuracy. The self-attention module in ClothFormer is incorporated with a lightweight backbone to achieve striking metrics on UTFPR-SBD3 and CFPD datasets.