Image and scene modelling in hazy conditions

Abstract

Outdoor scenes present specific challenges and opportunities for image processing and graphics, due to the presence of haze and other atmospheric effects. In this work, we present three major contributions, all of which involve hazy images. Firstly, we develop a flexible 3D haze model, based on systems of dynamic Gaussian blobs, which can deform and interact with the scene structure. The synthetic haze can be rendered in real-time, by interpolating its density along each optical ray. A GPU implementation is developed, using programmable shaders, and incorporated into an interactive HTML5 browser interface. The second major contribution of the thesis is an evaluation methodology for single image dehazing algorithms. The test data is generated by adding synthetic homogeneous and heterogeneous haze to real images. Physical consistency is achieved by combining the Gaussian haze model with high resolution lidar scans of each scene. This results in a realistic and challenging data set, which includes ground truth colour images, and associated optical transmission maps. A selection of current dehazing algorithms are evaluated on the new data set. The third major contribution is a novel approach to the construction of layered scene models, based on a single view of a hazy landscape. The underlying radiance image and transmission map are first estimated, by standard dehazing methods. The radiance image is then segmented into a small number of clusters, and a corresponding scene-plane is estimated for each cluster. The depth ordering of the planes is estimated from the optical transmission map, based on the principle that more distant parts of the scene were more hazy, in the original image. This approach provides the basic structure of a 2.5D scene model, without the need for multiple views, or image correspondences. The final models, which resemble cardboard ‘pop-ups’, are visually convincing, and can be manipulated interactively.