Locality-Aware Crowd Counting.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Crowd counting datasets often exhibit imbalanced data distributions, leading to significant under- and over-estimation issues.
• Existing crowd counting methods have not adequately addressed the challenges posed by data imbalance.

Purpose of the Study:

• To propose a novel locality-based learning paradigm to mitigate sample bias and produce generalizable features for crowd counting.
• To enhance the performance of existing deep crowd counting approaches by alleviating data imbalance problems.

Main Methods:

• Introduced a locality-aware data partition (LADP) approach using locality-sensitive hashing to group training data into balanced batches.
• Developed a locality-aware data augmentation (LADA) method that adaptively augments image patches based on loss to reduce training bias.
• Integrated the proposed methods with existing backbone network architectures in an end-to-end paradigm.

Main Results:

• The proposed locality-based learning paradigm effectively alleviates sample bias in crowd counting.
• Experiments demonstrate significant improvements in accuracy and generalization capabilities compared to state-of-the-art methods.
• The method showed versatility when applied to adversarial defense tasks.

Conclusions:

• The proposed locality-aware learning paradigm offers a simple yet effective solution for imbalanced data in crowd counting.
• The LADP and LADA methods can be seamlessly integrated with various deep learning architectures to boost performance.
• This approach advances crowd counting research by addressing a critical data imbalance challenge.