Reproduce “BEVANet: Bilateral Efficient Visual Attention Network for Real-Time Semantic Segmentation” 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
Ping-Mao Huang, I-Tien Chao, Ping-Chia Huang, Jia-Wei Liao, Yung-Yu Chuang
BEVANet is a real-time semantic segmentation network that utilizes a bilateral architecture, an efficient large kernel attention mechanism, and boundary-guided adaptive fusion. The model achieved an mIoU of 81.0% at 32.8 FPS on the Cityscapes validation set, meeting real-time requirements.
Key insights
- Decomposing large kernels into smaller convolutions and strip dilation kernels (SDLSKA) efficiently expands the receptive field for semantic information capture without excessive computational cost.
- Dynamically adjusting and fusing multi-scale features through a comprehensive channel-wise and spatial-wise selection mechanism (CKS) improves context understanding and detail refinement.
- Explicitly guiding the fusion of semantic and spatial features with boundary information (BGAF) enhances the accuracy of object contour delineation in segmentation masks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.