Research map / Image generation / Image restoration
CrowdDiff: Multi-hypothesis Crowd Density Estimation using Diffusion Models
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.
Frames crowd counting as conditional denoising diffusion that generates narrow-kernel density maps, counts by thresholding and fusing multiple sampled realizations, with an auxiliary training-only regression branch.
Problem
Density-map crowd counting suffers from background noise accumulation and density loss when broad Gaussian kernels are used, while narrow kernels perform poorly with regression networks.
Approach
Treat density map prediction as a conditional diffusion generation task so narrow kernels can be learned; count by thresholding density maps into dot maps rather than summing; fuse multiple stochastic realizations via SSIM-ordered, radius-based point fusion; and add a training-only count regression branch on U-Net features.
Key results
- Achieves MAE 47.3 on JHU-Crowd++, 47.4 on ShanghaiTech A, 5.7 on ShanghaiTech B, and 57.8 on NWPU-Crowd, reported as state-of-the-art in Table 1.
- Multi-realization fusion with four realizations lowers UCF-QNRF MAE from 74.59 (1 realization) to 68.95 and F-score rises from 67.89 to 81.48.
- Density thresholding beats density summation on JHU-Crowd++ (MAE 48.24 vs 215.40) for the same density maps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CrowdDiff instead of the strongest reported baseline | Crowd counting methods · JHU-Crowd++ | MAE: 52.3 -> 47.3 | Helped |
| Use CrowdDiff instead of the strongest reported baseline | Crowd counting methods · ShanghaiTech A | MAE: 49.3 -> 47.4 | Helped |
| Use CrowdDiff instead of the strongest reported baseline | Crowd counting methods · ShanghaiTech B | MAE: 5.7 -> 5.7 | No clear effect |
| Use CrowdDiff instead of the strongest reported baseline | Crowd counting methods · UCF-CC-50 | MAE: 184.1 -> 160.8 | Helped |
| Use CrowdDiff instead of the strongest reported baseline | Crowd counting methods · UCF-QNRF | MAE: 74.3 -> 68.9 | Helped |
| Use CrowdDiff instead of the strongest reported baseline | Crowd counting methods · NWPU-Crowd | MAE: 68.7 -> 57.8 | Helped |
| Add the counting decoder during training (ablation) | CrowdDiff · JHU-Crowd++ | MAE: 52.29 -> 48.56 | Helped |
| Add the counting decoder during training (ablation) | CrowdDiff · UCF-QNRF | MAE: 78.35 -> 73.08 | Helped |
| Fuse maps in ascending rather than descending SSIM order (ablation) | CrowdDiff · UCF-QNRF | MAE: 71.73 -> 68.95 | Helped |
| Increase the number of realizations from 1 to 4 (ablation) | CrowdDiff · UCF-QNRF | MAE: 74.59 -> 68.95 | Helped |
| Increase the number of realizations from 4 to 8 (ablation) | CrowdDiff · UCF-QNRF | MAE: 68.95 -> 66.97 | Helped |
| Use a 3×3 rather than a 5×5 or 9×9 density kernel (ablation) | CrowdDiff · UCF-QNRF | MAE: 81.87 -> 68.95; 94.68 -> 68.95 | Helped |
Surprising findings
- Broad kernels (5x5, sigma 1 and 9x9, sigma 2) degrade performance sharply compared with the 3x3 narrow kernel, with 1x1 performing similarly to 3x3.
Limitations
- Iterative diffusion sampling makes inference slower than regression-based counters.
- Fusing multiple realizations can introduce false positives and double counts when dots are shifted between realizations.
- Narrow kernels with broad-kernel-based baselines show performance sensitivity to kernel size and the rejection radius beta.
Experimental setup
Datasets: JHU-Crowd++ ShanghaiTech A ShanghaiTech B UCF-CC-50 UCF-QNRF NWPU-Crowd
Benchmarks: JHU-Crowd++ ShanghaiTech A ShanghaiTech B UCF-CC-50 UCF-QNRF NWPU-Crowd
Metrics: MAE MSE NAE precision recall F-score
Baselines: TopoCount SUA ChfL MAN GauNet CLTR CrowdHat STEERER PET DM-Count TransCrowd MATT
Builds on: DDPM DDIM Improved DDPM hybrid loss Min-SNR weighting P2PNet
Closest papers by meaning
- Distribution Matching for Crowd Counting (NeurIPS 2020 · similarity 0.89)
- Modeling Noisy Annotations for Crowd Counting (NeurIPS 2020 · similarity 0.89)
- DecideNet: Counting Varying Density Crowds Through Attention Guided Detection and Density Estimation (CVPR 2018 · similarity 0.87)
- Crowd Counting in the Frequency Domain (CVPR 2022 · similarity 0.87)
- Residual Regression With Semantic Prior for Crowd Counting (CVPR 2019 · similarity 0.87)
- Gaussian Density Splatting Network (NeurIPS 2026 · similarity 0.85)
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.