OBTAINING ROBUST GEOMETRIC AND PHOTOMETRIC INFORMATION FROM NIGHTTIME IMAGES

Abstract

The goal of this thesis is to extract robust geometric and photometric information from nighttime images.

In our first work, we estimate depth from nighttime stereo images based on joint structure-stereo optimization. We compute depth from smooth structure images obtained from the nighttime stereo images, constrained with a stereo-consistency relationship. We improve this work in our second work by proposing an unsupervised deep-learning method. We employ a network that performs unpaired day/night image translation and depth estimation jointly. In addition, we handle the fake depth problem which occurs due to unpaired image translation, for glow/glare and uninformative regions.

In our third work, we propose a non-deep-learning method to estimate the Camera Response Function (CRF) for general daytime and nighttime images. Our method is based on the ideas of prediction consistency and gradual refinement. Our method can be employed for nighttime enhancement that allows us to better recover the object surface colors. We improve this work in our fourth work by proposing a semi-supervised deep-learning method to enhance the visibility of the input nighttime image by estimating the CRF to increase the dynamic range of the intensity (to boost the intensity of the low-light regions), and simultaneously, suppress glow/glare light effects.