Benchmarking of Semantic Segmentation Enabled Human Activity Recognition Methods

Human part segmentation is a sub-part of semantic segmentation where it demands accurate and efficient image segmentation mechanisms for assessing the visual contents and real-time decision-making. It is a method for grouping together elements in an image that fall into the same class. In the case of humans, it is utilized to distinguish between their limbs and other body parts such as their elbows, shoulders, hands. The primary applications are scene understanding, navigation, video surveillance, virtual, augmented reality, and activity recognition. This paper aims to present an efficient comparison among various state-of-the-art methodologies for human body part segmentation. The underlying dataset was LIP and Pascal Person Part. In general, all the methodologies were found to perform better on Pascal Person Part dataset even though it contains only three thousand images in comparison to the LIP dataset, which consists of fifty thousand images, owing to better annotation and masking in the former. While the mean IoU achieved on the LIP dataset varied from 37.60 to 57.03%, it ranged between 58.49 to 78.61% for the Pascal person part dataset.