DTS-Net: Depth-to-Space Networks for Fast and Accurate Semantic Object Segmentation

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Semantic segmentation is crucial for scene understanding.
• Existing methods often rely on complex encoder-decoder architectures.
• Efficient real-time segmentation remains a challenge.

Purpose of the Study:

• To propose an effective and efficient technique for semantic segmentation.
• To introduce a novel network architecture inspired by depth-to-space image reconstruction.
• To explore joint semantic segmentation and depth estimation.

Main Methods:

• Developed Depth-to-Space Net (DTS-Net) and DTS-Net-Lite using efficient sub-pixel convolutions.
• Employed depth-wise separable convolution-based architectures (Xception, MobileNetV2).
• Integrated Nearest Label Filtration for mask enhancement.

Main Results:

• Achieved high mean intersection over union (mIOU) and mean pixel accuracy (Pix.acc.) on PASCAL VOC2012, NYUV2, and CITYSCAPES.
• Demonstrated superior performance compared to state-of-the-art methods, including those with encoder-decoder architectures.
• Showcased efficient real-time semantic segmentation capabilities with DTS-Net-Lite.

Conclusions:

• DTS-Net offers a simpler yet highly effective approach to semantic segmentation.
• The proposed method achieves competitive results with reduced computational complexity.
• The technique shows promise for joint semantic segmentation and depth estimation tasks.