Adopting ConvNeXt and contextual representation for enhanced feature integration in compact CPM for 2D hand pose estimation

Abstract

Hand Pose Estimation (HPE) is essential in Computer Vision (CV) applications such as Human-Computer Interaction (HCI), playing a critical role in fields like Virtual Reality (VR), robotics, medicine, and more. HPE, on the other hand, is confronted with obstacles like variances in hand size and the agility of hand movements. While 3D HPE has improved a lot, their dependency on 2D key points has led to a greater focus on enhancing 2D HPE methods. Deep learning has made significant advancements in these methods, especially in models including Deep Convolutional Neural Networks (DCNN) and Convolutional Pose Machines (CPM). In this study, we proposed a lightweight CPM for accurate 2D HPE to minimize the complexity of the model. The approach utilizes a modified ConvNeXt, incorporating a Global Context Block (GCB) as a central component. This integration is key for understanding and extracting enhanced features effectively. Our developed method demonstrates a notable improvement in performance, achieving an average accuracy increase of 2.62% compared to the Optimized Convolutional Pose Machine (OCPM), a state-of-the-art (SOTA) lightweight model in 2D HPE.