LSDNet: lightweight stochastic depth network for human pose estimation

Human pose estimation plays a critical role in human-centred vision applications. Its influence extends to various aspects of daily life, from healthcare diagnostics and sports training to augmented reality experiences and gesture-controlled interfaces. While current approaches have achieved impressive accuracy, their high model complexity and slow detection speeds significantly limit their deployment on edge devices with limited computing power, such as mobile phones and IoT devices. In this paper, we introduce a novel lightweight network for 2D human pose estimation, called lightweight stochastic depth network (LSDNet). Our approach is based on the observation that the majority of HRNet's parameters are located in the middle and later stages in the network. We reduce some unnecessary branches to significantly reduce these parameters. This is achieved by leveraging the Bernoulli distribution to randomly remove these redundant branches, which improves the network's efficiency while also increasing its robustness. To further reduce the network's parameter count, we introduce two lightweight blocks with simple yet effective architectures. These blocks achieve significant parameter reduction while maintaining good accuracy. Furthermore, we leverage coordinate attention to effectively fuse features from different branches and scales. This mechanism captures both inter-channel dependencies and spatial context, enabling the network to accurately localize keypoints across the human body. We evaluated the effectiveness of our method on the MPII and COCO datasets, demonstrating superior results on human pose estimation compared to popular lightweight networks. Our code is available at: https://github.com/illusory2333/LSDNet.