AAC: Automatic Augmentation for Crowd Counting

Abstract

Intelligent crowd management is important for city monitoring. In the task of crowd counting, insufficient training samples with labels are great challenges. In order to improve the performance of the counting model, data augmentation is an effective method. This paper proposes an automatic augmentation framework for counting (AAC) based on the deep reinforcement learning. We first pre-train the model and iteratively generate a data augmentation strategy based on the Deep Deterministic Policy Gradient (DDPG) algorithm on the divided validation dataset. When the optimal augmentation action for the model on a specific dataset is found, the model is fine-tuned by the augmented dataset. At the same time, we release a large-scale crowd counting dataset HaCrowd describing Hajj scenario. Finally, five typical crowd counting models are used to carry out augmentation experiments on four small datasets including HaCrowd. The experimental results show that AAC method can be used to generate flexible augmentation strategy for counting model aimed for specific datasets. With the augmentation dataset to train model, it can further improve the performance of counting model. The download link of the HaCrowd is (https://github.com/KAU-Smart-Crowd/HaCrowd).