Reproduce “JetSeg: Efficient Real-Time Semantic Segmentation Model for Low-Power GPU-Embedded Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Miguel Lopez-Montiel, Daniel Alejandro Lopez, Oscar Montiel
JETSEG, developed by researchers at CITEDI-IPN, introduces an encoder-decoder model for real-time semantic segmentation tailored for low-power GPU-embedded systems, integrating novel architectural components and a specialized loss function. The model achieves 158 FPS on an NVIDIA TITAN RTX GPU and 39.9 FPS on an NVIDIA Jetson AGX, demonstrating efficient performance with only 0.03M parameters and 1.125 GFLOPs.
Key insights
- The JetConv operator effectively enlarges the receptive field and captures long-range dependencies through combined convolutions while maintaining a low parameter count.
- Efficient architectural units like JetBlock, incorporating group convolutions and lightweight activation functions (TanhExp), optimize the balance between inference time, memory usage, and model abstraction.
- A tailored loss function (JetLoss) with adaptive class weighting significantly improves segmentation accuracy and training efficiency, especially for underrepresented or challenging classes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.