Multi-Granularity Distribution Alignment for Cross-Domain Crowd Counting.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) transfers knowledge from labeled source to unlabeled target domains.
• UDA is increasingly applied to crowd counting, but global distribution alignment struggles with intra-domain variations (density, angle, scale).
• Existing methods face challenges with inaccurate alignment and computational inefficiency due to intra-domain gaps.

Purpose of the Study:

• To propose a novel Multi-Granularity Optimal Transport (MGOT) framework for UDA in crowd counting.
• To address fine-grained, domain-agnostic variations within crowd counting datasets.
• To improve the accuracy and efficiency of crowd counting models in cross-domain scenarios.

Main Methods:

• Developed the Multi-Granularity Optimal Transport (MGOT) framework for distribution alignment.
• Phase 1: Clustering coarse-grained features based on intra-domain similarity.
• Phase 2: Aligning granular clusters using optimal transport and mapping cluster centers to patch levels.
• Phase 3: Re-weighting the aligned distribution for model refinement in domain adaptation.

Main Results:

• MGOT framework demonstrates superior performance in adaptive crowd counting.
• Achieved state-of-the-art results across twelve cross-domain benchmarks.
• Outperformed existing methods in handling intra-domain disparities.

Conclusions:

• MGOT effectively aligns domain-agnostic factors at multiple granularities, overcoming limitations of global alignment.
• The proposed method enhances the robustness and accuracy of crowd counting models in diverse, unlabeled target domains.
• This work offers a significant advancement in unsupervised domain adaptation for crowd counting applications.