Reproduce “$CrowdDiff$: Multi-hypothesis Crowd Density Estimation using Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yasiru Ranasinghe, Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, Vishal M. Patel
Researchers at Johns Hopkins University introduce CrowdDiff, which applies denoising diffusion probabilistic models to crowd density estimation, enabling the use of narrow Gaussian kernels and a novel thresholding-based counting strategy to overcome issues of density loss and background noise. The system generates multiple density map hypotheses that are fused, achieving new state-of-the-art results across several public datasets, including an MAE of 57.8 and MSE of 221.2 on NWPU-Crowd.
Key insights
- Diffusion models can effectively learn from and reproduce density maps generated with narrow Gaussian kernels, overcoming limitations of prior regression methods and preserving distinct head representations.
- Counting by thresholding discrete kernels in density maps provides superior immunity to background noise and more precise localization compared to traditional density summation.
- Leveraging the inherent stochasticity of generative diffusion models through a systematic multi-hypothesis fusion process significantly improves the robustness and accuracy of crowd counting.
- Joint training with a weakly-supervised auxiliary counting branch guides the denoising network's feature learning, leading to more robust individual density map realizations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.