From a General-Purpose Object-Detector to an Efficient Multi -Human Parsing Network

Human parsing is a critical and valuable task, often utilized as an intermediate module in comprehensive image and video analysis systems. For instance, in human-centric applications, human parsing supports tasks such as re-identification, fashion change detection, and human attribute extraction. When analyzing images that include humans, cars, and motorbikes, human parsing specifically aids in the analysis of human-related aspects. However, the overall system must also detect and parse other objects, such as cars and motorbikes, alongside humans. Consequently, this paper aims to integrate human parsing into a general-purpose object detector. We selected Yolov8 [1], a widely-used technique in the industry, as the backbone for object detection and adapted it for this purpose. While our method is categorized as a multi-human parsing technique, the principles outlined can be extended to parsing other objects, such as cars and motorbikes. Our experimental comparisons show that our approach main-tains Yolov8's advantageous features while achieving performance comparable to state-of-the-art methods in multi-human parsing.