Accurate and Real-Time Variant Hand Pose Estimation Based on Gray Code Bounding Box Representation

Effective hand gestures are crucial for human-machine interactions, and recent advancements in neural networks offer promising avenues for efficient hand pose estimation. However, existing methods still face challenges in detecting hand poses of different scales within a single RGB image sensor. This article introduces a novel approach, drawing inspiration from modulus conversion, to enhance the efficiency of hand pose estimation from a single RGB image sensor. The method involves transforming the floating-point values of hand regions into binary codes, ensuring continuous numerical space without a significant computational overhead. This approach significantly improves accuracy for hands of varying sizes in both detection and pose estimation. Furthermore, this article addresses the challenge of datasets lacking hand keypoints annotations by introducing a novel loss computation for labeled keypoints during network training. To assess the effectiveness of the proposed method, a new benchmark for variant hand scales is presented, facilitating evaluation across different hand sizes. The proposed approach undergoes testing on diverse datasets, with experimental results demonstrating comparable performance to state-of-the-art methods, thereby validating its efficacy. Additionally, the study conducts several ablation studies, exploring aspects such as the choice of Gray code, code length, effectiveness across different hand scales, and training with labeled keypoints to affirm the efficiency and effectiveness of the proposed method.