Task-Oriented Feature-Fused Network With Multivariate Dataset for Joint Face Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep multitask learning is gaining traction for face analysis, but existing methods struggle with task-specific feature requirements, varying convergence rates, and limited annotated data.
• Optimizing a primary task while learning auxiliary tasks presents challenges in balancing performance across different facial analysis objectives.

Purpose of the Study:

• To develop an effective multitask framework for simultaneous face detection, landmark localization, and attribute analysis.
• To address limitations in current multitask learning approaches for face analysis, including feature representation and training optimization.

Main Methods:

• Propose a Task-Oriented Feature-Fused Network (TFN) incorporating a task-oriented feature-fused block for learning task-specific feature combinations.
• Introduce an alternative multitask training scheme to optimize individual task convergence rates within the joint framework.
• Present the Jilin Face Analysis (JFA) dataset, a large-scale dataset with multivariate annotations for face detection, landmark localization, and attribute analysis.

Main Results:

• The TFN demonstrates superior performance compared to existing multitask models on the JFA dataset.
• Achieves competitive results on benchmark datasets like WIDER FACE and 300W for face detection and landmark localization.
• Obtains state-of-the-art performance in gender recognition on the MORPH II dataset.

Conclusions:

• The proposed TFN effectively handles the complexities of multitask learning for face analysis by learning task-specific features and optimizing training.
• The JFA dataset provides valuable resources for advancing research in comprehensive face analysis.
• The TFN framework offers a promising direction for improving performance in various face-related computer vision tasks.