Crowd Counting via Joint SASNet and a Guided Batch Normalization Network

Abstract

Recent studies on crowd counting have achieved promising results by using convolutional neural network (CNN) architectures. However, due to the large variation in scene distribution in real-world crowd datasets, it remains a challenge to achieve high performance using standard CNN methods. Such methods often suffer from performance drops. To address this challenge, this paper proposes a new crowd-counting approach that combines three state-of-the-art methods: Guided-Batch-Normalization, which adapts the model using unseen dataset normalization parameters; the Scale Adaptive Selection Network, which uses a multi-level network to obtain variation feature representations; and Distribution-Matching-Count, which uses a new loss function between prediction and ground truth maps. Combining these methods results in improved performance. Extensive experiments across multiple datasets have demonstrated that the proposed approach outperforms state-of-the-art methods.