ConvNeXtPose: A Fast Accurate Method for 3D Human Pose Estimation and Its AR Fitness Application in Mobile Devices

초록

In general, 3D human-pose estimation requires high-performance computing resources. Existing methods working on mobile devices trade off accuracy in return for increased efficiency, often making the estimation accuracy far from sufficient for developing serious applications. In this paper, we present a mobile 3D human-pose estimation model, achieving real-time performances with a well-designed balance between efficiency and accuracy. As the backbone, our model leverages the cutting-edge ConvNeXt architecture, renowned for its feature extraction capabilities. We enhance its performance through strategic architectural modifications and incorporation of depthwise separable convolutions in the upsampling module. The experiments made with the Human3.6M dataset show that the accuracy delivered by our model is comparable to that of the state-of-the-art models, consuming significantly fewer computational resources. To showcase the practicality of our model, we present a prototype of an AR fitness application. Built upon our 3D human pose estimation model, it helps trainees recreate trainers' poses from reference images. The effectiveness of the application is validated via experiments and evaluations. The source code can be found at: https://github.com/medialab-ku/ConvNeXtPose.

키워드