Low-Resolution Self-Attention for Semantic Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Semantic segmentation requires high-resolution details for pixel accuracy and global context for class prediction.
• Existing vision transformers face computational bottlenecks due to high-resolution context modeling.

Purpose of the Study:

• To introduce a computationally efficient mechanism for capturing global context in semantic segmentation.
• To develop a vision transformer that addresses the computational limitations of current models.

Main Methods:

• Developed the Low-Resolution Self-Attention (LRSA) mechanism for global context modeling.
• Computed self-attention in a fixed low-resolution space, augmented with depth-wise convolutions for high-resolution details.
• Built the LRFormer, an encoder-decoder vision transformer incorporating the LRSA mechanism.

Main Results:

• The LRFormer model demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant reductions in computational cost (FLOPs) through the LRSA mechanism.
• Validated effectiveness across diverse datasets: ADE20K, COCO-Stuff, and Cityscapes.

Conclusions:

• The LRSA mechanism offers an effective and efficient approach to semantic segmentation.
• LRFormer presents a promising alternative to existing vision transformers for high-performance, low-computation tasks.
• The proposed method advances the field of efficient deep learning for computer vision.