Abstarct: Instance segmentation is an important task in Computer vision which has many applications in industries and real life. One important task is instance segmentation for autonomous vehicles and traffic control in urban scenes. With recent developments in AI and especially deep learning, many solutions were proposed to address this task. Our goal is to improve accuracy of existing instance segmentation methods. One important challenge in supervised training is that they need large amounts of labeled data which is expensive to annotate. It is possible to generate new data baxsed on existing ones. We propose a novel frxamework to copy instances of humans from a dataset image and paste it on another one. It has an adversarial network baxsed on deep reinforcement learning which decides where to paste the human instance. An encoder network using A2C algorithm suggests positions and a critic decides how real the picture looks like after placement. After predicting the place to paste, we have to put the object instance over the background in a realistic way. This is done with another adversarial network which learns to seamlessly blend the object with the background. With this approach we could synthesize new data and increase the mAP of instance segmentation by 2.1%.