Research map / Label-efficient and robust learning / Semi-supervised learning
Taste More, Taste Better: Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting
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.
Proposes TMTB, a semi-supervised crowd counting framework combining diffusion-based background inpainting augmentation, a VMamba (visual state space) backbone, and an anti-noise classification head within a Mean Teacher setup.
Problem
Semi-supervised crowd counting struggles to effectively use unlabeled images because standard augmentations (Mixup, CutMix) destroy the spatial structure of crowd density maps, and CNN backbones overfit local details and noisy point annotations hurt regression.
Approach
Inpaint only background regions with Stable Diffusion to diversify unlabeled crowd images while preserving foreground people, weight the inpainted-sample loss by a classification-branch inconsistency mask, and use a VMamba backbone with an auxiliary count-bin classification head that gives less exact but more accurate supervision.
Key results
- With 5% labels on JHU-Crowd++, TMTB reaches MAE 67.0 (RMSE 252.2), versus 76.5 for prior SOTA MRC-Crowd.
- With 5% labels on UCF-QNRF, MAE 96.3 vs 101.4 for MRC-Crowd; with 10% on ShanghaiTech A, MAE 65.7 vs 67.3 for MRC-Crowd.
- Ablation on UCF-QNRF (5%) shows VSSM with classification and inpainting loss progressively reduce MAE from 115.35 to 96.33.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TMTB instead of the strongest listed baseline | Semi-supervised crowd counting · JHU-Crowd++, 5% labeled | MAE: 76.5 -> 67.0 | Helped |
| Use TMTB instead of the strongest listed baseline | Semi-supervised crowd counting · UCF-QNRF, 5% labeled | MAE: 101.4 -> 96.3 | Helped |
| Use TMTB instead of the strongest listed baseline | Semi-supervised crowd counting · ShanghaiTech B, 10% labeled | MAE: 9.7 -> 9.8 | Hurt |
| Add VSSM to the supervised regression setup (ablation) | TMTB · UCF-QNRF, 5% labeled | MAE: 115.35 -> 110.41 | Helped |
| Add the classification loss (ablation) | VSSM + supervised regression · UCF-QNRF, 5% labeled | MAE: 110.41 -> 105.88 | Helped |
| Add the semi-supervised loss (ablation) | VSSM + regression + classification · UCF-QNRF, 5% labeled | MAE: 105.88 -> 103.58 | Helped |
| Add inpainting augmentation with fixed inconsistency-level weights (ablation) | VSSM + regression + classification + semi-supervised loss · UCF-QNRF, 5% labeled | MAE: 103.58 -> 97.21 | Helped |
| Use time-based rather than fixed inpainting weights (ablation) | TMTB · UCF-QNRF, 5% labeled | MAE: 97.21 -> 96.33 | Helped |
| Apply Mixup to the whole image rather than use TMTB inpainting augmentation | Crowd counting augmentation · ShanghaiTech A, 10% labeled | MAE: 65.74 -> 285.16 | Hurt |
| Apply Mixup to the background only rather than use TMTB inpainting augmentation | Crowd counting augmentation · ShanghaiTech A, 10% labeled | MAE: 65.74 -> 72.00 | Hurt |
| Apply CutMix rather than use TMTB inpainting augmentation | Crowd counting augmentation · ShanghaiTech A, 10% labeled | MAE: 65.74 -> 312.60 | Hurt |
| Apply BCP rather than use TMTB inpainting augmentation | Crowd counting augmentation · ShanghaiTech A, 10% labeled | MAE: 65.74 -> 70.99 | Hurt |
Surprising findings
- Mixup and CutMix applied to whole images degrade performance heavily (MAE 285.16 and 312.60 vs 65.74 for TMTB on ShanghaiTech A, 10%).
Limitations
- Inpainted images can contain noisy or fragmented regions, requiring an inconsistency-based weighting mask.
- Method relies on pseudo-labels with no confidence estimation for regression maps.
Experimental setup
Datasets: JHU-Crowd++ UCF-QNRF ShanghaiTech A ShanghaiTech B
Benchmarks: JHU-Crowd++ UCF-QNRF ShanghaiTech A ShanghaiTech B
Metrics: MAE RMSE
Baselines: MT L2R DACount OT-M MRC-Crowd CU SALCrowd MTCP SUA Mixup CutMix BCP
Builds on: Mean Teacher VMamba MRC-Crowd BCP CUT
Closest papers by meaning
- CrowdCLIP: Unsupervised Crowd Counting via Vision-Language Model (CVPR 2023 · similarity 0.87)
- Leveraging Unlabeled Data for Crowd Counting by Learning to Rank (CVPR 2018 · similarity 0.87)
- Free Lunch Enhancements for Multi-modal Crowd Counting (CVPR 2025 · similarity 0.86)
- CrowdDiff: Multi-hypothesis Crowd Density Estimation using Diffusion Models (CVPR 2024 · similarity 0.85)
- Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting (CVPR 2025 · similarity 0.84)
- Single Domain Generalization for Crowd Counting (CVPR 2024 · similarity 0.84)
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.