Unsupervised Depth Estimation for Night Images Using CycleGAN

Depth estimation is the method of estimating the distance between a camera and the objects in a given image/scenario. We can also say, that it's a method of estimating a three-dimensional representation of a two-dimensional image. The depth map that will be generated using this method will infer information about the distance of each point from the camera. Depth estimation has wide applications, in various fields like robotics, autonomous vehicles, virtual reality, and augmented reality. Let's propose an example of autonomous vehicles which use depth estimation to detect obstacles and avoid collisions. The CycleGAN model can also be deployed to estimate the depth of the given set of images. In this model, the generator will be fed with RGB images as an input along with a random noise vector, and as a result, it will produce a depth map as output. The conditional input, in this case, could be any kind of auxiliary information that helps to constrain the generation process, such as the camera parameters or the surface normals of the scene. The novelty of the present paper is that we are trying to convert daylight images into night images using CycleGAN and feeding the night images as input into another CycleGAN model to estimate the depth of the night images. The present work has also estimated the depth of day images using CycleGAN. The proposed model has been trained and tested on two benchmark datasets.