Domain-Adaptive Crowd Counting via High-Quality Image Translation and Density Reconstruction

Recently, crowd counting using supervised learning achieves a remarkable improvement. Nevertheless, most counters rely on a large amount of manually labeled data. With the release of synthetic crowd data, a potential alternative is transferring knowledge from them to real data without any manual label. However, there is no method to effectively suppress domain gaps and output elaborate density maps during the transferring. To remedy the above problems, this article proposes a domain-adaptive crowd counting (DACC) framework, which consists of a high-quality image translation and density map reconstruction. To be specific, the former focuses on translating synthetic data to realistic images, which prompts the translation quality by segregating domain-shared/independent features and designing content-aware consistency loss. The latter aims at generating pseudo labels on real scenes to improve the prediction quality. Next, we retrain a final counter using these pseudo labels. Adaptation experiments on six real-world datasets demonstrate that the proposed method outperforms the state-of-the-art methods.

I. Introduction

Crowd counting is usually treated as a pixel-level estimation problem, which predicts the density value for each pixel and sums the entire prediction map as a final counting result. A pixelwise density map produces more detailed information than a single number for a complex crowd scene. In addition, it also boosts other highly semantic crowd analysis (group detection [1]–[3], crowd segmentation [4], public management [5], and so on) or video surveillance tasks (video summarization [6]–[8] and abnormal detection [9]). Recently, benefiting from the powerful capacity of deep learning, there is a significant promotion in the field of counting. However, currently released datasets are too small to satisfy the mainstream deep learning-based methods [10]–[15]. The main reason is that constructing a large-scale crowd counting dataset is extremely demanding, which needs many human resources [16].