Self-Correction for Human Parsing

Labeling pixel-level masks for fine-grained semantic segmentation tasks, e.g.human parsing, remains a challenging task. The ambiguous boundary betweendifferent semantic parts and those categories with similar appearance usuallyare confusing, leading to unexpected noises in ground truth masks. To tacklethe problem of learning with label noises, this work introduces a purificationstrategy, called Self-Correction for Human Parsing (SCHP), to progressivelypromote the reliability of the supervised labels as well as the learned models.In particular, starting from a model trained with inaccurate annotations asinitialization, we design a cyclically learning scheduler to infer morereliable pseudo-masks by iteratively aggregating the current learned model withthe former optimal one in an online manner. Besides, those correspondinglycorrected labels can in turn to further boost the model performance. In thisway, the models and the labels will reciprocally become more robust andaccurate during the self-correction learning cycles. Benefiting from thesuperiority of SCHP, we achieve the best performance on two popularsingle-person human parsing benchmarks, including LIP and Pascal-Person-Partdatasets. Our overall system ranks 1st in CVPR2019 LIP Challenge. Code isavailable at https://github.com/PeikeLi/Self-Correction-Human-Parsing.