Research map / Detection and segmentation / Open-set and few-shot detection
MaSS13K: A Matting-level Semantic Segmentation Benchmark
Read the paperCheck your idea against 108,967 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.
Introduces MaSS13K, a 13,348-image 4K matting-level semantic segmentation dataset, and MaSSFormer, a Mask2Former-based model with a high-resolution pixel decoder, benchmarked against 14 segmentation methods.
Problem
Existing semantic segmentation datasets have low resolution and coarse masks, and standard segmentation models struggle to produce fine boundaries and remain efficient at high resolution.
Approach
Builds a large 4K dataset with matting-level semantic masks across seven categories, and a Mask2Former-based pixel decoder that fuses global semantic features with edge-guided low-level detail features; also uses pseudo-labels for new classes with label re-weighting.
Key results
- MaSSFormer (R50) reaches 88.97 mIoU, 48.97 BIoU and 0.5639 BF1 on MaSS-val, the best among 15 compared methods.
- On MaSS-test, MaSSFormer reaches 88.21 mIoU and 48.39 BIoU, versus 88.00 mIoU and 46.13 BIoU for Mask2Former.
- MaSSFormer-Lite (ResNet-18) reaches 87.11 mIoU on MaSS-val with 15.07M parameters and 771G FLOPs at 4096x4096.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MaSSFormer instead of the strongest ResNet-50 baseline, Mask2Former | MaSSFormer · MaSS13K-val | mIoU: 88.28 -> 88.97 | Helped |
| Use MaSSFormer instead of MPFormer | MaSS13K-val · MaSS13K-val | BIoU: 47.81 -> 48.97 | Helped |
| Add CST for global semantic aggregation (ablation) | MaSSFormer · MaSS13K-val | mIoU: 85.54 -> 88.11 | Helped |
| Add RFB to broaden the receptive field (ablation) | MaSSFormer with CST · MaSS13K-val | mIoU: 88.11 -> 88.29 | Helped |
| Add LSE to extract low-level structure features (ablation) | MaSSFormer with CST and RFB · MaSS13K-val | BIoU: 45.23 -> 47.36 | Helped |
| Add LSE to extract low-level structure features (ablation) | MaSSFormer with CST and RFB · MaSS13K-val | mIoU: 88.29 -> 88.02 | Hurt |
| Add EGF for edge-guided fusion (ablation) | MaSSFormer with CST, RFB, and LSE · MaSS13K-val | mIoU: 88.02 -> 88.97 | Helped |
| Add EGF for edge-guided fusion (ablation) | MaSSFormer with CST, RFB, and LSE · MaSS13K-val | BIoU: 47.36 -> 48.97 | Helped |
| Add precise MaSS13K labels to pseudo-label training (ablation) | MaSSFormer on Car · Car | BIoU: 20.44 -> 22.52 | Helped |
| Apply label re-weighting to reduce the weight of mislabeled pseudo-label edges (ablation) | MaSSFormer on Car · Car | mIoU: 83.23 -> 95.21 | Helped |
| Apply label re-weighting to reduce the weight of mislabeled pseudo-label edges (ablation) | MaSSFormer on Car · Car | BIoU: 31.81 -> 35.68 | Helped |
Surprising findings
- Training directly on pseudo-labels for a new class did not improve segmentation quality, and mixing incorrect edge pseudo-labels caused a 10.95% mIoU drop on the Car class without re-weighting.
Limitations
- Mask complexity and resolution advantages are shown only against the chosen datasets and metrics; benefit for downstream tasks is shown only by a single bokeh example.
- New-class transfer is demonstrated only for one class (Car) with pseudo-labels from existing models.
Experimental setup
Datasets: MaSS13K ADE20K
Benchmarks: MaSS13K validation MaSS13K test
Metrics: mIoU BIoU BF1
Baselines: STDC2 BiSeNetV2 SegNeXt PIDNet FeedFormer SeaFormer CGRSeg DeepLabv3+ UperNet OCRNet MaskFormer Mask2Former
Builds on: Mask2Former MaskFormer SE block Deformable convolution
Closest papers by meaning
- Boosting Semantic Human Matting With Coarse Annotations (CVPR 2020 · similarity 0.89)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.87)
- Integrating Low-Level Visual Cues for Enhanced Unsupervised Semantic Segmentation (AAAI 2025 · similarity 0.87)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.87)
- $\mathbf{D^{3}S^{2}}$: Diffusion-Guided Dataset Distillation for Semantic Segmentation (NeurIPS 2026 · similarity 0.86)
- Bridge the Points: Graph-based Few-shot Segment Anything Semantically (NeurIPS 2024 · 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.