Learning Reasoning-Decision Networks for Robust Face Alignment.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional face alignment methods often suffer from biased predictions due to poor initialization in coarse-to-fine approaches.
• Unconstrained environments present challenges like pose deformations and appearance variations, complicating accurate face alignment.

Purpose of the Study:

• To propose an end-to-end Reasoning-Decision Network (RDN) approach for robust face alignment.
• To overcome limitations of conventional methods by learning a robust initialization policy directly from raw pixels.
• To improve face alignment accuracy and robustness in challenging, unconstrained scenarios.

Main Methods:

• Formulated face alignment as a Markov decision process, defining an agent to learn an optimal shape searching policy.
• Leveraged raw pixels to reason shape candidates and sequentially made decisions to remove outliers for robust initialization.
• Developed a Long Short-Term Memory (LSTM)-based value function to evaluate shape quality and adjusted gradients using policy gradients to avoid local optima.

Main Results:

• The proposed Reasoning-Decision Network (RDN) approach demonstrates robust face alignment capabilities.
• RDN consistently outperforms most state-of-the-art face alignment methods.
• The approach showed superior performance on four widely-evaluated challenging datasets.

Conclusions:

• The end-to-end RDN approach offers a significant advancement in robust face alignment.
• Policy gradient-based learning effectively addresses initialization bias and improves performance in complex environments.
• RDN provides a more accurate and reliable solution for face alignment tasks, especially under challenging conditions.