Domain Adaptation in Crowd Counting

We consider the problem of domain adaptation in crowd counting. Given an input image of a crowd scene, our goal is to estimate the count of people in the image. Previous work in crowd counting usually assumes that training and test images are captured by the same camera. We argue that this is not realistic in real-world applications of crowd counting. In this paper, we consider a domain adaptation setting in crowd counting where we have a source domain and a target domain. For example, these two domains might correspond to cameras at two different locations (i.e., with differing viewpoints, illumination conditions, environment objects, crowd densities, etc.). We have enough labeled training data from the source domain, but we only have either unlabeled data or a small number of labeled data in the target domain. Our goal is to train a crowd counting system that performs well in the target domain. We believe this setting is closer to real-world deployment of crowd counting systems. Due to the domain shift, a model trained from the source domain is unlikely to perform well in the target domain. In this paper, we propose several domain adaptation techniques for this problem. Our experimental results demonstrate the superior performance of our proposed approach on several benchmark datasets.