Unsupervised Domain Adaption, which was recently developed, focuses on the adaptation of knowledge from labeled source data to unlabeled target data. Domain adaptation is an approach that efficiently adapts a new model to different environments where the source and target domains are similar but not identical. By adapting the information from the source domain to the target domain, it improves the probability of enhancing performance. However, the majority of the research were plagued by a domain shift issue. They neglect crucial areas that correspond to the image. To minimize the domain gap in unsupervised domain adaptation, we suggest a simple yet efficient spatial distribution-based Sum pooling approach and Consistency regularization in this paper. The sum pooling approach integrates information taken from the model in both the height and width axes. Consistency regularization avoids sum pooling divergence and tackles the domain shift problem by shortening the distance between the feature of encoder and the sum pooling feature. We evaluate our technique on datasets such as MNIST, MNIST-M, SVHN for classification tasks, and Cityscapes, Foggy Cityscape, Sim10k, and KITTI for object detection tasks. The results show that our proposed approach is helpful in decreasing the distance between the source and target domains.
