Joint Hand Detection and Rotation Estimation Using CNN.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Hand detection is crucial for tasks like pose recovery and gesture recognition.
• Uncontrolled environments present challenges due to hand flexibility and cluttered backgrounds.
• Existing methods struggle with in-plane rotations and diverse conditions.

Purpose of the Study:

• To develop a robust hand detection method for uncontrolled environments.
• To jointly address hand detection and in-plane rotation estimation.
• To improve the accuracy and reliability of hand detection systems.

Main Methods:

• A convolutional neural network (CNN) architecture based on Faster R-CNN.
• Explicit formulation of in-plane rotation within the network.
• A dedicated rotation network to estimate hand orientation.
• A derotation layer to align feature maps for improved detection.

Main Results:

• The proposed method achieves state-of-the-art performance on widely-used benchmarks (Oxford, Egohands).
• Joint rotation estimation and detection mutually benefit classification accuracy.
• Demonstrated superior performance compared to existing hand detection models.

Conclusions:

• The novel CNN effectively handles in-plane rotations for improved hand detection.
• The joint approach enhances robustness in complex, real-world scenarios.
• This method offers a significant advancement for hand-related computer vision tasks.