Virtual Classification: Modulating Domain-Specific Knowledge for Multidomain Crowd Counting.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multidomain crowd counting aims to develop generalizable models across diverse datasets.
• Deep learning models often exhibit domain bias, favoring dominant data distributions over others.

Purpose of the Study:

• To propose a novel network, MDKNet, to mitigate domain bias in multidomain crowd counting.
• To enable deep networks to effectively model varied distributions from multiple datasets with minimal bias.

Main Methods:

• Introduced the Modulating Domain-Specific Knowledge Network (MDKNet) utilizing a "modulating" concept.
• Developed an instance-specific batch normalization (IsBN) module to adapt information flow to domain distributions.
• Incorporated a domain-guided virtual classifier (DVC) to learn a domain-separable latent space for modulator guidance.

Main Results:

• MDKNet demonstrated superior performance in tackling multidomain crowd counting challenges.
• The proposed methods effectively balanced and modeled diverse dataset distributions.
• Experiments on Shanghai-tech A/B, QNRF, and NWPU benchmarks validated MDKNet's effectiveness.

Conclusions:

• MDKNet successfully addresses the domain bias issue in multidomain crowd counting.
• The proposed IsBN and DVC modules enhance model adaptability and generalization across datasets.
• The approach offers a significant advancement for learning from heterogeneous crowd counting data.