We consider a problem of domain adaptation in image classification and crowd counting. Given a pre-trained model learned from a source domain, our goal is to adapt this model to a target domain using unlabeled data. The solution of this problem has a lot of potential applications in computer vision research that require a neural network model adapted to a target dataset. In this thesis, we propose two different approaches for domain adaptation. First, inspired by a source free domain adaptation, we propose a black-box model adaptation and distillation for image classification. The key challenge of this problem setting is that we do not have access to any internal information of the source model, including model architecture, model parameters, or even intermediate feature maps. We can only access the output of the source model, hence the source model is a “black-box” to us. Once the model is adapted to the target domain, we perform knowledge distillation to obtain a compact model for deployment. Second, we apply dynamic transfer for solving domain adaptation problems in crowd counting. The key insight is that adapting the model for the target domain is achieved by adapting the model across the data samples. The experimental results on several benchmark datasets demonstrate the effectiveness of our approaches.