Reproduce “Fine-grained Domain Adaptive Crowd Counting via Point-derived Segmentation” on OpenResearch

About this paper

Yongtuo Liu, Dan Xu, Sucheng Ren, Hanjie Wu, Hongmin Cai, Shengfeng He

Researchers at South China University of Technology, University of Amsterdam, and Hong Kong University of Science and Technology developed a fine-grained domain adaptation framework for crowd counting, using weakly-supervised segmentation derived from sparse point annotations to guide targeted feature alignment and density regularization. The method achieved state-of-the-art performance, demonstrating, for example, a 30.7% MAE gain in Synthetic-to-Real adaptation (GCC SHPartB) compared to prior unsupervised approaches.

Key insights

• Fine-grained domain adaptation, by separating crowd and background features, is more effective than holistic approaches for crowd counting, as background variations often hinder robust crowd feature learning.
• Reliable crowd segmentation masks can be derived from sparse point-level annotations through a weakly-supervised approach, eliminating the need for costly pixel-level labeling.
• Targeted adversarial alignment within crowd regions (CRT) and regularization of target-domain density using its own estimated crowd distribution (CDA) significantly improve adaptation performance.

Summary by alphaXiv.