BMSeNet: Multiscale Context Pyramid Pooling and Spatial Detail Enhancement Network for Real-Time Semantic...

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Real-time semantic segmentation networks often use shallow architectures, limiting receptive fields and single-scale feature extraction.
• This leads to reduced generalization, robustness, and accuracy due to loss of spatial details.

Purpose of the Study:

• To propose a novel network, BMSeNet, that addresses limitations in real-time semantic segmentation.
• To improve segmentation accuracy, robustness, and generalization by incorporating multiscale context and spatial detail enhancement.

Main Methods:

• Introduced a Multiscale Context Pyramid Pooling Module (MSCPPM) to aggregate multiscale contextual information and enlarge receptive fields.
• Designed a Spatial Detail Enhancement Module (SDEM) to compensate for lost spatial details and improve perception.
• Proposed a Bilateral Attention Fusion Module (BAFM) to effectively merge features from different branches using pixel positional correlations.

Main Results:

• BMSeNet demonstrated a strong balance between inference speed and segmentation accuracy.
• The proposed network outperformed several state-of-the-art real-time semantic segmentation methods on benchmark datasets.
• Experimental validation was conducted on the Cityscapes and CamVid datasets.

Conclusions:

• BMSeNet effectively addresses the limitations of shallow architectures in real-time semantic segmentation.
• The integration of MSCPPM, SDEM, and BAFM significantly enhances segmentation performance.
• The proposed network offers a promising solution for accurate and robust real-time semantic segmentation.