Small Object Augmentation of Urban Scenes for Real-Time Semantic Segmentation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Deep learning significantly improved semantic segmentation accuracy for autonomous driving.
• High-quality models like PSPNet and DeepLabV3 are computationally expensive, hindering real-time applications.
• Existing real-time methods struggle with segmenting small, critical objects like traffic lights.

Purpose of the Study:

• To enhance the performance of real-time semantic segmentation for autonomous driving.
• To develop a computationally efficient yet accurate segmentation model.
• To improve the detection of small objects crucial for safety.

Main Methods:

• Proposed a novel real-time semantic segmentation model named Narrow Deep Network (NDNet).
• Developed a synthetic dataset by augmenting training images with additional small objects.
• Evaluated NDNet's performance and re-trained existing models (PSPNet, DeepLabV3) on the new dataset.

Main Results:

• NDNet achieved 65.7% mean intersection over union (mIoU) on the Cityscapes test set with low computational cost (8.4G FLOPs).
• Re-training PSPNet and DeepLabV3 on the synthetic dataset resulted in an average 2% mIoU improvement on small objects.
• The proposed methods address the trade-off between real-time performance and accuracy in semantic segmentation.

Conclusions:

• NDNet offers an efficient solution for real-time semantic segmentation in autonomous driving.
• The synthetic dataset effectively improves small object segmentation for existing models.
• The research contributes to safer and more practical autonomous driving systems through improved scene understanding.