Usually,we chose train and test datasets from one distribution.It means that for making a dataset after gathering enough photos,we use most of the images for training and keep some photos for test.When we train a network and test it with similar data of training networks,it works very well during training.But after that,when we want to use this network in reality,the efficiency decreases a lot.This is a challenge that domain adaptive deal with it and tries to solve.The challenge is that although we know where we want to use our network and have an image of the target area,these data are not annotated.Making labels for these data is time-consuming and expensive,so we need to have another solution for this problem which domain adaptive is an area for solving it.We approach this task by dividing our development into two parts,first improve the image to image generation with one of the newest networks in Generative Adversarial Network（GAN）area and increase the accuracy of object detection with reduce wrong datasets with filter and improve the backbone based on dense network.Here we tried that with Generative Adversarial Network,and Siamese Network makes a robust network for generating new photos.It means that with the label and content of the old dataset,it makes a new dataset whose style is similar to the target area.And then,with a filter,we reduced the number of defective labels,finally trained the detection module with these labels with a robust backbone.Our method show increased performance in object detection with the new dataset.The result was successful for cityscapes,foggy-cityscapes,and the GTA5 dataset