Reproduce “Cross-Image Relational Knowledge Distillation for 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
Chuanguang Yang, Helong Zhou, Zhulin An, Xue Jiang, Yongjun Xu, Qian Zhang
A framework for semantic segmentation, Cross-Image Relational Knowledge Distillation (CIRKD), transfers global semantic pixel relationships across entire image datasets from a teacher to a student network. It improved the mean Intersection-over-Union (mIoU) of lightweight student models by up to 0.78% on Cityscapes and showed consistent gains on other benchmarks like CamVid and Pascal VOC.
Key insights
- Global semantic relations among pixels, extending beyond individual image boundaries, constitute valuable knowledge for enhancing the performance of student networks in dense prediction tasks.
- A well-trained teacher network develops a robust pixel embedding space that inherently captures more accurate and extensive pixel correlations, including those spanning different images.
- Memory-bank mechanisms, adapted from self-supervised learning, can effectively mitigate the limitations of small batch sizes in semantic segmentation by facilitating the capture and distillation of broad, long-range pixel dependencies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.