Surface reflection model estimation from naturally illuminated image sequences

This thesis addresses the problem of estimating the surface reflection model of objects observed in a terrestrial scene, illuminated by natural illumination; that is, a scene which is illuminated by sun and sky light alone. This is a departure from the traditional analysis of laboratory scenes, which are illuminated by idealised light sources with positions and radiance distributions that are precisely controlled. Natural illumination presents a complex hemispherical light source which changes in both spatial and spectral distribution with time, terrestrial location, and atmospheric conditions.

An image-based approach to the measurement of surface reflection is presented. The use of a sequence of images, taken over a period of time, allows the varying reflection from the scene due to the changing natural illumination to be measured. It is shown that the temporal change in image pixel values is suitable for the parameters of a reflection model to be estimated. These parameters are estimated using regression techniques. Two such regression methods are considered: a traditional non-linear method and the probabilistic approach of simulated annealing. It is shown that simulated annealing provides consistent performance in this application.

This work focuses on the use of physically-based models of illumination, surface reflection and camera response. Using such models allows the system to produce quantitative, as opposed to qualitative, results and allows radiometric measurements to be made from image pixel values. The use of accurate models of daylight illumination allows scenes illuminated by skies of varying atmospheric conditions to be considered. The results obtained by the presented methods may be used for a variety of tasks ranging from object recognition to the automated generation of virtual environments.

Results are presented which show that the proposed method is suitable for the wide variety of camera positions, surface orientations and sky conditions that may be experienced. The method is also shown to be tolerant of image noise and may be used on single or multiple pixels within each image. These results are based on the analysis of synthetic image sequences generated using a validated lighting simulation system. Results are also presented for real data recorded using a camera.