SD-HRNet: Slimming and Distilling High-Resolution Network for Efficient Face Alignment

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Face alignment is crucial for high-level face analysis tasks like human activity recognition.
• Existing facial landmark detection models are often parameter-heavy and computationally inefficient for practical use.

Purpose of the Study:

• To develop a lightweight facial landmark detector.
• To improve computational efficiency and reduce model parameters for practical applications.

Main Methods:

• Proposing a network-level architecture-slimming method for high-resolution supernetworks.
• Introducing a selective feature fusion mechanism to prune redundant operations.
• Implementing a triple knowledge distillation scheme with peer student networks and a teacher network.

Main Results:

• Achieved competitive performance on benchmarks like 300W, COFW, and WFLW.
• Demonstrated a favorable trade-off between model parameters (0.98 M-1.32 M) and FLOPs (0.59 G-0.6 G).
• Outperformed recent state-of-the-art methods in terms of efficiency.

Conclusions:

• The proposed architecture-slimming and knowledge distillation methods result in an efficient lightweight facial landmark detector.
• The approach offers a practical solution for real-time face analysis applications requiring high accuracy and low computational cost.