Study On Reflectance Reconstruction Techniques In Reverse Rendering

In this thesis, we address the reflectance reconstruction of objects from images. Reflectance properties, illumination conditions and geometries of the scene play important roles in human perception of real world objects. Reflectance reconstruction from photographs in order to re-render high realistic images is in the scope of inverse rendering technique, and its application in virtual reality and augmented reality has a very good prospect. Because of the variety of objects, no method could be applied to all kinds of objects. Therefore, we have to divide and conquer each kind of them. In addition, the specularity and texture of real world objects drive reflectance reconstruction to a much harder way. Focusing those difficult issues of reflectance recovery, we conduct some research, and our major contributions include the following aspects:Firstly, an effective method is proposed to reconstruct the reflectance properties of objects with specularity. In order to approach an accurate description of the reflectance, there are benefits to separate diffuse/specular reflectance component. In our method, specularity detection and separation algorithm are applied to reliably separate the diffuse/specular reflectance component. Then, the separation results are used to independently fit parameters of reflectance model. Compared with methods which using measurements to directly fit parameters of reflectance model, our method is more accurate and faster.Secondly, a method is presented to recover the reflectance of objects with multi-materials. For a textured object, it is very difficult to recover reflectance, since each point on the surface may have different reflectance properties. We assume the object is composed with several basic materials, and each point in the surface can be represented as a weighted sum of them. Therefore, the problem of reflectance recovery is converted to determine BRDF (Bidirectional Reflection Distribution Function) parameters of each basic material and weight for each point. First, a clustering algorithm is applied to automatically compute the number of basic materials, and an initial weight map is estimated based on the clustering results. Then, the weight map and parameters of the reflectance model are iteratively optimized. In this step, diffuse and specular albedo is estimated separately to exhibit more reliable results. Finally, the reflectance could be reconstructed under arbitrary illumination/viewing conditions using our output parameters. Besides those objects composed with distinct materials, this method is also useful to reconstruct objects mixed with complex materials, such as alloy and marble. For the latter, no special sampling device is needed, and the reflectance can be reliably recovered with a few images.Experiments on synthetic and real scenes were conducted to validate our approaches.