Research map / Detection and segmentation / Open-set and few-shot detection
FreeMask: Synthetic Images with Dense Annotations Make Stronger Segmentation Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FreeMask synthetic images during joint training | Mask2Former + Swin-T · ADE20K | mIoU: 48.7 -> 52.0 | Helped |
| add FreeMask synthetic images during joint training | Mask2Former + Swin-T · COCO-Stuff | mIoU: 44.5 -> 46.4 | Helped |
| pre-train on FreeMask synthetic images, then fine-tune on real images | SegFormer + MiT-B4 · ADE20K | mIoU: 48.5 -> 50.6 | Helped |
| add pixel-level filtering of noisy synthetic regions (ablation) | SegFormer + MiT-B4 trained on synthetic images · ADE20K synthetic-to-real transfer | mIoU: 43.3 -> 47.4 | Helped |
| add hardness-aware re-sampling of masks (ablation) | SegFormer + MiT-B4 trained on synthetic images · ADE20K synthetic-to-real transfer | mIoU: 43.3 -> 45.1 | Helped |
| apply both pixel filtering and hardness-aware re-sampling (ablation) | SegFormer + MiT-B4 trained on synthetic images · ADE20K synthetic-to-real transfer | mIoU: 43.3 -> 48.3 | Helped |
| add synthetic images without filtering or re-sampling (ablation) | real-image baseline · ADE20K joint training | mIoU: 48.5 -> 48.2 | Hurt |
| add filtering and re-sampling to synthetic images in joint training (ablation) | SegFormer + MiT-B4 real-image baseline · ADE20K | mIoU: 48.5 -> 50.3 | Helped |
| double real-only training iterations (ablation) | real-image baseline · ADE20K | mIoU: 48.7 -> 49.2; 48.5 -> 48.6; 46.2 -> 46.7 | No clear effect |
| increase Nmax from 6 to 20 with filtering and re-sampling (ablation) | SegFormer + MiT-B4 synthetic-only training · ADE20K | mIoU: 47.2 -> 48.3 | Helped |
| use hardness-aware re-sampling instead of OHEM (ablation) | Segmenter-S · ADE20K | mIoU: 44.2 -> 45.4 | Helped |
Closest papers by meaning
- Network-Free, Unsupervised Semantic Segmentation With Synthetic Images (CVPR 2023 · similarity 0.89)
- MaskFactory: Towards High-quality Synthetic Data Generation for Dichotomous Image Segmentation (NeurIPS 2024 · similarity 0.89)
- A Training-free Synthetic Data Selection Method for Semantic Segmentation (AAAI 2025 · similarity 0.89)
- MaSS13K: A Matting-level Semantic Segmentation Benchmark (CVPR 2025 · similarity 0.86)
- A Generative Approach at the Instance-Level for Image Segmentation Under Limited Training Data Conditions (Student Abstract) (AAAI 2025 · similarity 0.86)
- $\mathbf{D^{3}S^{2}}$: Diffusion-Guided Dataset Distillation for Semantic Segmentation (NeurIPS 2026 · 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.