Rethinking Portrait Matting with Privacy Preserving

Area of Science:

• Computer Vision
• Machine Learning
• Privacy-Preserving AI

Background:

• Traditional portrait matting relies on identifiable images, posing privacy risks.
• Existing methods lack robust privacy-preserving capabilities.
• Need for anonymized datasets to train and evaluate privacy-preserving models.

Purpose of the Study:

• Introduce P3M-10k, a large-scale anonymized benchmark for Privacy-Preserving Portrait Matting (P3M).
• Develop and evaluate a unified matting model (P3M-Net) and a novel strategy (P3M-CP) for enhanced privacy and generalization.
• Address the cross-domain performance gap in privacy-preserving training (PPT).

Main Methods:

• Created P3M-10k dataset: 10,421 high-resolution, face-blurred portraits with alpha mattes.
• Developed P3M-Net, a unified matting model adaptable to CNN and transformer architectures.
• Implemented P3M-CP strategy to reacquire facial context using public celebrity images at data and feature levels.

Main Results:

• P3M-Net demonstrates superiority over state-of-the-art methods on P3M-10k and public benchmarks.
• P3M-CP significantly improves cross-domain generalization ability under the privacy-preserving training setting.
• Systematic evaluation of trimap-free and trimap-based matting methods under PPT.

Conclusions:

• P3M-10k serves as a crucial benchmark for privacy-preserving portrait matting research.
• P3M-Net and P3M-CP offer effective solutions for privacy concerns and generalization in matting.
• The findings hold significant implications for future research and real-world applications of privacy-preserving AI.