We consider the problem of scene adaptive crowd counting. Given a target camera scene, our goal is to adapt a model to this specific scene with only a few labeled/unlabeled images. The solution to this problem has potential applications in numerous real-world scenarios that require deploying a crowd counting model specially adapted to a target camera. In this thesis, we propose two novel methods for scene adaptive crowd counting. First, inspired by the recently introduced learning to learn paradigm in the context of few-shot regime, we aim to learn the parameters of a crowd counting model in a way to facilitate fast adaptation to the target scene. Second, we introduce a new problem called unlabeled scene adaptive crowd counting. More specifically, we propose to use few unlabeled images from the target scene to perform the adaptation. We introduce a novel AdaCrowd framework to solve this problem and it consists of a crowd counting network and a guiding network. The guiding network predicts some parameters in the crowd counting network based on the unlabeled images from a particular scene. This allows our model to adapt to different target scenes. The experimental results on several challenging benchmark datasets demonstrate the effectiveness of our two proposed approaches.