Home       Publications      York       Polarisation      Photometric stereo      Q-Bot      Composites


Work from the University of York

I am a former PhD student of the University of York, Department of Computer Science. I worked under the supervision of Edwin Hancock in the Computer Vision and Pattern Recognition group.

My thesis, Surface Shape and Reflectance Analysis Using Polarisation, was accepted in May 2007. My broad area of research was on the spontaneous polarisation of light caused by surface reflection, and the information contained in this polarisation state. In particular, I was able to combine information from two views using patch matching techniques to acquire stereo correspondence, which I then used to enhance the shape recovery process. I also experimented with combining polarisation information with shape-from-shading. Details of my thesis are given below.


Here are some examples of reconstructions using the technique described in the PAMI paper, November 2007.




Summary of Thesis


When unpolarised light is reflected from a smooth dielectric surface, it becomes partially polarised. This thesis aims to exploit the polarising properties of surface reflection for computer vision applications. Most importantly, the thesis proposes novel shape and reflectance function estimation techniques. The methods presented rely on polarisation data acquired using a standard digital camera and a linear polariser. Fresnel Theory lies at the heart of the thesis and is used to process the polarisation data in order to estimate the surface normals of the target objects.

The first novel technique is a simple single-view approach to shape reconstruction, where efficiency is given priority over accuracy. The technique, along with the precision of the underlying theory, is extensively tested and compared to ground truth.

More accurate shape estimation techniques are later presented that incorporate shading information into the surface normal estimation process. This is achieved using robust statistics to estimate how the measured pixel brightnesses depend on the surface orientation. This gives an estimate of the object material reflectance function, which is then used to refine the estimates of the surface normals. The techniques use the histograms of surface orientation angles and pixel intensities and fit reflectance functions to the data using probability density functions and simulated annealing.

The most sophisticated and accurate shape reconstruction algorithm presented uses the refined normal estimates to establish a two-view correspondence. To do this, a set of patches are extracted from each view and are aligned by minimizing an energy functional based on the surface normal estimates and local topographic properties. The optimum alignment parameters for different patch pairs are then used to establish stereo correspondence. Our techniques are most suited to smooth, non-metallic surfaces. The multi-view method complements existing stereo algorithms since it does not require salient surface features to obtain correspondences.