AnyFace++: Deep Multi-Task, Multi-Domain Learning for Efficient Face AI.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Accurate face detection and facial landmark localization are crucial for applications like emotion recognition, age estimation, and gender identification.
• Current deep learning models often require multiple specialized models for different tasks, leading to increased memory usage and inference time.
• Existing research primarily focuses on human faces, neglecting other domains like animals and cartoon characters.

Purpose of the Study:

• To develop an input-agnostic face model capable of performing multiple face-related tasks concurrently.
• To extend face analysis capabilities beyond human faces to include animal and cartoon domains.
• To address the limitations of current models regarding memory usage and inference time.

Main Methods:

• Proposed AnyFace++, a deep multi-task, multi-domain learning model.
• Utilized a heterogeneous cost function for training.
• Trained the model on diverse datasets encompassing human, animal, and cartoon faces.

Main Results:

• AnyFace++ successfully performs face detection and facial landmark prediction for human, animal, and cartoon faces.
• The model achieves comparable performance to state-of-the-art models specialized for specific domains.
• Enabled concurrent execution of tasks such as age estimation, gender classification, and emotion recognition for human faces.

Conclusions:

• AnyFace++ offers a unified and efficient solution for various face-related computer vision tasks across multiple domains.
• The model's input-agnostic nature and multi-task capability significantly reduce computational overhead.
• This research advances the field by providing a versatile tool for broader facial analysis applications.