Robust object counting through distribution uncertainty matching and optimal transport.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Object counting is often framed as density estimation from point-annotated images.
• Point-based annotation is cost-effective but susceptible to noise, impacting model performance.

Purpose of the Study:

• To develop a robust object counting method that addresses annotation noise.
• To introduce a novel loss function, DUMLO (Distribution Uncertainty Matching for Loss Optimization), for density estimation.

Main Methods:

• DUMLO models uncertainty over augmented points to define a loss function between ground-truth and target density maps.
• The loss function is formulated as a coupling of two optimal transport problems.
• A new algorithm, Trihorn, is proposed to jointly estimate the loss function and the augmentation set's density map, quantifying annotation uncertainty.

Main Results:

• Theoretical analysis confirms a tight generalization error bound for the proposed loss function.
• Extensive evaluation on pathology, crowd, and vehicle datasets demonstrates strong performance.
• The model achieves good Mean Absolute Error and shows robustness to annotation noise.

Conclusions:

• DUMLO offers a robust and accurate approach to object counting, particularly in the presence of noisy annotations.
• The method exhibits fast convergence properties.
• The Trihorn algorithm effectively quantifies annotation uncertainty, contributing to improved density estimation models.