EvHandPose: Event-Based 3D Hand Pose Estimation With Sparse Supervision.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Event cameras offer low-power, high-dynamic-range imaging suitable for 3D hand pose estimation.
• Challenges include motion ambiguity and difficulties in annotating dense event streams.

Purpose of the Study:

• To develop an accurate and robust event-based 3D hand pose estimation method.
• To address motion ambiguity and sparse annotation issues in event data.
• To create a large-scale, real-world dataset for event-based hand pose estimation.

Main Methods:

• Proposed EvHandPose with novel hand flow representations in the Event-to-Pose module.
• Introduced contrast maximization and hand-edge constraints in the Pose-to-IWE module for weak supervision.
• Developed EvRealHands, a large-scale real-world dataset for event-based hand pose.

Main Results:

• EvHandPose outperformed previous event-based methods on the EvRealHands dataset.
• Achieved accurate, stable, high-temporal-resolution hand pose estimation in challenging conditions (fast motion, strong light).
• Demonstrated strong generalization to outdoor scenes and different event camera types.

Conclusions:

• EvHandPose effectively alleviates motion ambiguity and performs robustly under sparse annotation.
• The method achieves state-of-the-art performance for event-based hand pose estimation.
• EvRealHands dataset facilitates future research and bridges the real-synthetic domain gap.