Reproduce “S$^2$-RL: Sample-Set Dual Reinforcement Learning for Generative Semantic Segmentation Dataset Distillation” 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
Haoyu Wang, Fei Zhou, Qingqing Qiu, Lei Zhang, Wei Wei, Chen Ding
S²-RL constructs compact segmentation training sets by fine-tuning a text-to-image diffusion model with feedback on individual images and semantic coverage across the set. It uses category prompts and prototypes informed by real data. At a 0.25% ADE20K budget, SegFormer-B5 reaches 15.60% mIoU, versus 13.54% for pretrained SD3.5. Results favor S²-RL across tested budgets on ADE20K and COCO, but evidence is limited to these benchmarks and tested students.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.