Person Foreground Segmentation by Learning Multi-Domain Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Accurate single person segmentation is crucial for human-related research and photo editing.
• Current segmentation algorithms lack either precision or real-time performance.

Purpose of the Study:

• To develop a novel real-time single person image segmentation model.
• To address the limitations of existing general or slow segmentation methods.

Main Methods:

• Introduced a multi-domain learning framework within the Multi-domain TriSeNet Networks.
• Utilized a Feature Fusion Module (FFM) to create domain-independent and domain-specific layers.
• Employed a self-supervised learning strategy to enhance domain knowledge transfer and accuracy.

Main Results:

• The Multi-domain TriSeNet achieved state-of-the-art performance on public and custom datasets.
• The model demonstrated real-time processing speeds.
• A new large-scale dataset, MSSP20k, was created for single person image segmentation.

Conclusions:

• The proposed Multi-domain TriSeNet effectively performs real-time single person image segmentation.
• The multi-domain learning approach and self-supervised strategy significantly improve accuracy.
• The MSSP20k dataset provides a valuable resource for advancing segmentation research.