Neuron Linear Transformation: Modeling the Domain Shift for Crowd Counting.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Cross-domain crowd counting (CDCC) is crucial for public safety, aiming to reduce domain shifts between datasets.
• Existing methods often use image translation or adversarial learning to find domain-invariant features.
• Domain shifts are also observable at the model parameter level, presenting an alternative adaptation approach.

Purpose of the Study:

• To propose a novel method, Neuron Linear Transformation (NLT), for addressing domain shifts in crowd counting at the parameter level.
• To directly model the domain gap by learning domain shift parameters for individual neurons.

Main Methods:

• The proposed Neuron Linear Transformation (NLT) method learns domain shift parameters using a few labeled target data points.
• NLT exploits domain factors and bias weights to define the transformation for source model neurons.
• A linear transformation is applied to generate target model neurons, adapting the model to the target domain.

Main Results:

• NLT achieved state-of-the-art performance on six real-world datasets for cross-domain crowd counting.
• The method demonstrated superior effectiveness compared to supervised training and fine-tuning approaches.
• Ablation studies confirmed the robustness and high efficacy of the NLT method.

Conclusions:

• Neuron Linear Transformation (NLT) offers a powerful and effective approach for cross-domain crowd counting by directly addressing parameter-level domain shifts.
• The method shows significant improvements over existing domain adaptation techniques.
• NLT provides a robust and accurate solution for crowd counting tasks in diverse environments.