Reproduce “AICSD: Adaptive Inter-Class Similarity 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
Amir M. Mansourian, Rozhan Ahmadi, Shohreh Kasaei
Researchers at Sharif University of Technology introduce Adaptive Inter-Class Similarity Distillation (AICSD), a framework that enhances lightweight semantic segmentation models by transferring high-order inter-class relationships and dynamically adjusting teacher supervision. AICSD improved mean Intersection-over-Union (mIoU) by up to 4.37% on Cityscapes and 4.13% on Pascal VOC 2012 compared to baseline student models.
Key insights
- Distilling explicit high-order inter-class similarity provides richer structural guidance than conventional pixel-wise or channel-wise methods, improving a student network's understanding of semantic boundaries and relationships.
- Dynamically adjusting the weight of teacher supervision allows the student to leverage teacher expertise early in training while gradually relying more on ground truth and high-level structural guidance, mitigating negative impacts of teacher errors.
- The combination of structured relational knowledge transfer and adaptive supervision effectively bridges the performance gap between complex teacher models and lightweight student models for dense prediction tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.