Heatmap-Guided Selective Feature Attention for Robust Cascaded Face Alignment.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face alignment is crucial for facial landmark detection, commonly addressed by coordinate and heatmap regression.
• Existing multi-task learning networks struggle with training these tasks simultaneously due to differing feature map requirements and shared noisy features.

Purpose of the Study:

• To propose an efficient multi-task learning network for robust cascaded face alignment.
• To improve face alignment performance by effectively training coordinate and heatmap regression tasks.

Main Methods:

• Developed a heatmap-guided selective feature attention mechanism for multi-task learning.
• Implemented a background propagation connection to enhance feature map validity for each task.
• Employed a cascaded refinement strategy using heatmap regression for global landmarks and coordinate regression for precise localization.

Main Results:

• The proposed network demonstrated superior performance in face alignment tasks.
• Achieved state-of-the-art results on benchmark datasets including 300W, AFLW, COFW, and WFLW.
• Effectively addressed the challenge of training coordinate and heatmap regression tasks simultaneously.

Conclusions:

• The proposed heatmap-guided selective feature attention network offers a robust and efficient solution for cascaded face alignment.
• This approach enhances facial landmark detection accuracy by optimizing multi-task learning for regression tasks.
• The method sets a new standard for face alignment performance on various challenging datasets.