Hybrid Perturbation Strategy for Semi-Supervised Crowd Counting.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Conventional Convolutional Neural Network (CNN)-based crowd counting methods are susceptible to texture perturbations and imperceptible noises from adversarial attacks.
• Effective utilization of unlabeled data remains a challenge in semi-supervised crowd counting.

Purpose of the Study:

• To propose a simple yet effective semi-supervised crowd counting method based on consistency regularization.
• To enhance information mining from unlabeled images using a hybrid perturbation strategy.
• To improve the robustness of crowd counting models against various perturbations.

Main Methods:

• A hybrid perturbation strategy combining spatial texture transformation and adversarial perturbation modules is employed to perturb unlabeled data in semantic and non-semantic spaces.
• A cross-distribution normalization technique is introduced to stabilize model optimization under strong perturbations, addressing issues with Batch Normalization (BN) layers.
• The method leverages consistency regularization for semi-supervised learning in crowd counting.

Main Results:

• The proposed semi-supervised crowd counting method outperforms state-of-the-art methods on benchmark datasets including ShanghaiTech, UCF-QNRF, NWPU-Crowd, and JHU-Crowd++.
• The method demonstrates superior robustness against diverse perturbations compared to existing approaches.
• Extensive experiments validate the effectiveness and stability of the proposed approach.

Conclusions:

• The developed semi-supervised crowd counting method offers a significant advancement over current techniques.
• The hybrid perturbation strategy and cross-distribution normalization effectively enhance model performance and robustness.
• This work provides a robust and efficient solution for crowd counting using limited labeled data.