Research map / Detection and segmentation / Open-set and few-shot detection
S$^2$-RL: Sample-Set Dual Reinforcement Learning for Generative Semantic Segmentation Dataset Distillation
Haoyu Wang, Fei Zhou, Qingqing Qiu, Lei Zhang, Wei Wei, Chen Ding
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes S²-RL, a sample-set dual reinforcement learning framework for generative dataset distillation targeting semantic segmentation.
Problem
Dataset distillation for semantic segmentation: bi-level optimization is computationally expensive due to pixel-wise gradient unrolling, while proxy-based generative methods only add textures to fixed masks derived from real data.
Approach
Uses dual reinforcement learning over both individual samples and the sample set to jointly generate image-mask pairs for distillation, rather than texturing fixed masks; details beyond the truncated abstract are unknown.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic alignment reward (R_align) to GRPO fine-tuning of SD3.5 (ablation) | S2-RL · ADE20K, 1.0% budget | segmentation performance: 21.66 -> 22.91 | Improved |
| add intra-class diversity reward (R_div) on top of alignment (ablation) | S2-RL · ADE20K, 1.0% budget | segmentation performance: 22.91 -> 24.73 | Improved |
| add set-level reward (R_set) on top of alignment and diversity (ablation) | S2-RL · ADE20K, 1.0% budget | segmentation performance: 24.73 -> 25.47 | Improved |
| add set-level reward (R_set) on top of alignment and diversity (ablation) | S2-RL · ADE20K, 0.5% budget | segmentation performance: 20.83 -> 21.52 | Improved |
| add set-level reward (R_set) on top of alignment and diversity (ablation) | S2-RL · ADE20K, 0.25% budget | segmentation performance: 15.17 -> 15.60 | Improved |
| full dual-reward RL fine-tuning of SD3.5 versus vanilla SD3.5 (ablation) | SD3.5 Medium · ADE20K, 0.25% budget | segmentation performance: 13.54 -> 15.60 | Improved |
| add set-level reward to broaden coverage | S2-RL · ADE20K 1.0% layout diagnostics | geometry coverage: .549 -> .693 (vs SD3.5) | Improved |
| broaden generated category combinations via set-level reward | S2-RL · ADE20K 1.0% category-pair | pair precision: .955 -> .898 (UniDD* vs S2-RL) | Hurt |
| distill with S2-RL to different student architectures (ResNet-18 to 101, ConvNeXt, Swin, DeepLabV3, SegFormer) | S2-RL distilled datasets · ADE20K student variation | segmentation performance | Improved |
Closest papers by meaning
- $\mathbf{D^{3}S^{2}}$: Diffusion-Guided Dataset Distillation for Semantic Segmentation (NeurIPS 2026 · similarity 0.89)
- SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models (NeurIPS 2026 · similarity 0.87)
- T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · similarity 0.87)
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.86)
- Seg4Diff: Unveiling Open-Vocabulary Semantic Segmentation in Text-to-Image Diffusion Transformers (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
