SPPNet: Single-Person Human Parsing and Pose Estimation in RGB Videos

The human-centric visual analysis field thrives on rich video datasets that explore human behaviours and interactions. Yet, a gap persists in datasets covering both human pose estimation and parsing challenges. In this study, a notable effort has been made to develop a dedicated dataset named “Single Person Video-in-Person (SP-VIP)” to suit the research scenario, resolving a lack of a universal dataset to support three major human-centric visual analysis methods. The SP-VIP dataset was derived by extracting videos from the VIP dataset initially designed exclusively for parsing-related tasks. Furthermore, the VIP dataset did not encompass provisions for pose estimation and human activity recognition, which are crucial elements for human activity recognition. To bridge this gap, the SP-VIP dataset was meticulously curated with a specific focus on single-person activities. Videos in the newly created dataset are split into frames with semantic labels and joint values for each frame. To assess the performance of the tailored dataset, a novel architecture Single-person Parsing and Pose Network (SPPNet) was employed using a Deep ConvNet network for parsing while simultaneously performing pose estimation using the stacked hourglass method. To demonstrate the effectiveness of the newly created dataset, extensive experiments were performed on the discussed architecture, which produced favourable results with a pixel accuracy of 88.50%, a mean accuracy of 60.50%, and a mean Intersection over Union (IoU) of 49.30% signifying enhancement in performance.