Double Similarity Distillation for Semantic Image Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Achieving high accuracy and speed in semantic image segmentation is challenging, especially for compact networks with limited resources.
• Existing compact segmentation networks often have constrained performance due to resource limitations.

Purpose of the Study:

• To propose a novel knowledge distillation framework, Double Similarity Distillation (DSD), to improve the accuracy of compact semantic segmentation networks.
• To capture similarity knowledge in both pixel and category dimensions to enhance network performance.

Main Methods:

• Introduced a pixel-wise similarity distillation (PSD) module using residual attention maps for detailed spatial dependency capture.
• Developed a category-wise similarity distillation (CSD) module to strengthen global category correlation via a correlation matrix.
• Integrated PSD and CSD into the DSD framework with no extra parameters and minimal increase in Floating Point Operations (FLOPs).

Main Results:

• The DSD framework demonstrated significant improvements in classification accuracy across various compact networks.
• Experiments on Cityscapes, CamVid, ADE20K, and Pascal VOC 2012 datasets showed DSD outperforming state-of-the-art methods.
• The PSD module reduced computation while enhancing spatial feature extraction, and CSD improved global category understanding.

Conclusions:

• The proposed Double Similarity Distillation (DSD) framework is effective and general for improving compact semantic segmentation networks.
• DSD offers a practical solution for balancing accuracy and speed in resource-constrained environments.
• The framework's ability to capture both pixel-level and category-level similarities contributes to its superior performance.