GAN BASED OBJECT DETECTION OF NOISY IMAGES

Abstract

Due to increased use of CGI imagery in many application across different fields, it is high time there is a Generative Adversarial Network which worked in synergy with an object detection algorithm like YOLO to overcome the difficulty in perception of various kinds of noises contributing to the decreased accuracy of object class prediction. The problem with CGI imagery in real time is that they are replete with objects which are of different proportions when compared to real life objects and also they anthropomorphize every type of object like cars, trees, houses, toys which makes it difficult for default anchor boxes to act as good priors in drawing the bounding boxes around them. So we have integrated a network which can generate denoised images with the help of generative and discriminator networks competing against each other and the generated denoised image will directly be pushed through another object detection network which here is YOLOv3 with improved IoU. So with the integration of these networks and tuning the parameters for our custom dataset the output will be an image which can be used for real-time rendering.