Reproduce “BiSeNet V2: Bilateral Network with Guided Aggregation 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

Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, Nong Sang

BiSeNet V2 introduces a bilateral architecture for real-time semantic segmentation, balancing spatial detail and semantic context through two parallel branches. The network achieved 72.6% mean IoU at 156 FPS on the Cityscapes dataset using an NVIDIA GTX 1080 Ti.

Key insights

  • Real-time semantic segmentation can benefit from a dedicated dual-branch architecture that processes spatial details and semantic context separately and efficiently.
  • A guided aggregation mechanism can effectively combine features from detail-rich and semantic-rich branches to yield high-quality predictions.
  • Auxiliary training heads can improve model performance by providing stronger supervision signals during training, without increasing inference latency.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.