Progressive Glass Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Glass segmentation is crucial for computer vision tasks but challenging due to glass's lack of unique visual appearance and complex backgrounds.
• Existing methods struggle with fusing features from different network levels due to characteristic gaps, leading to suboptimal performance.

Purpose of the Study:

• To develop an effective feature fusion strategy for accurate glass segmentation.
• To address the challenge of fusing multi-level features in deep networks for improved glass detection.

Main Methods:

• Proposed a Progressive Glass Segmentation Network (PGSNet) utilizing Discriminability Enhancement (DE) and Focus-and-Exploration Based Fusion (FEBF) modules.
• DE modules bridge the gap between level-specific features, enhancing their discriminability for fusion.
• FEBF modules highlight commonalities and explore differences between features for richer information excavation.

Main Results:

• PGSNet progressively aggregates features from high-level to low-level for coarse-to-fine segmentation.
• The method achieved state-of-the-art performance, outperforming 26 models on three datasets across four metrics.
• Introduced the first home-scene-oriented glass segmentation dataset to support household robot applications.

Conclusions:

• The proposed PGSNet effectively tackles the challenges of glass segmentation through innovative feature fusion techniques.
• The developed dataset and model offer significant advancements for household robotics and computer vision research.