CAN HUMAN ATTRIBUTE SEGMENTATION BE MORE ROBUST TO OPERATIONAL CONTEXTS WITHOUT NEW LABELS?

Human Attribute Segmentation (HAS) describes, pixelwise, the different semantic parts of people in an image. This fine-grained description is useful for several applications (e.g. security, fashion). However, despite the good performance reached by supervised Semantic Segmentation (SS) approaches, they are usually biased by the source training dataset and suffer from a performance drop when applied on new domains. Annotating pixelwise images of each new encountered context is tedious and expensive. How can HAS be more robust to new contexts without annotating new images? This work is a first study coping with Unsupervised Domain Adaptation (UDA) for HAS. We combine self-supervised and semi-supervised learning paradigms with HPTR, an efficient Human Parsing method, to evaluate the impact on performance. UDA-HPTR, the resulting method, improves performance on both target (unlabeled) and source (labeled) datasets. Besides, it outperforms state-of-the-art UDA method for SS, mostly known in autonomous driving benchmarks, while using only half the number of parameters.