We present a method for enhancing the visual quality of existing digital elevation models textured with orthophotos, using a sparse set of unordered, high resolution photographs. After an initial manual selection of correspondence points, we automatically register the input photographs to the given terrain data set using robust image-based modeling techniques. To combine the geo-registered images on the terrain surface, we propose a compositing algorithm that ensures smooth transitions between the images while at the same time preserving the fine details. The resulting textures are inserted into the quadtree representation of a terrain rendering engine to allow an efficient realtime visualization. We demonstrate our method on an HRSC terrain data set and a collection of high resolution photos of Turtmann valley in Switzerland.

We propose a three-scale hierarchical representation of scenes and objects and show how this representation is suitable for both capture of models from images and efficient photo-realistic rendering. The model consists of: (1) a conventional triangulated geometry on the macro-scale; (2) a displacement map, introducing pixel-wise depth with respect to each planar model facet (triangle) on the meso level; (3) a photo-realistic micro-structure represented by an appearance basis spanning viewpoint variation in texture space. We implement a capture and rendering system for this model. Conventional Shape-From-Silhouette or Structure-From-Motion is used to capture the coarse macro geometry, variational shape and reflectance estimation for the meso-level, and texture basis optimization for the micro level. For efficiency the meso and micro level routines are both HW accelerated. Photo-realistic capture of complex scenes is thus possible in a few minutes using budget cameras and PC's, and rendering is real-time. Experimental results and videos show models from regular images of humans and objects.

Estimating light sources from images is an important and difficult problem in computer vision. High quality lighting is useful as input to other computer vision algorithms and in graphics rendering. For instance, photometric stereo and shape from shading requires known light and are limited to highly controlled laboratory setups. Accurate lighting is also helpful in augmented reality in order to consistently relight an artificially introduced object. Algorithms using individual point lights are useful for simple lighting setups, but for complex illumination a basis representation is needed. We propose a lighting model using Daubechies wavelets and a method for recovering light from cast shadows and specular highlights in images. We assume that the geometry is known for part of the scene. We tested our method for difficult cases of both uniform and textured objects and under complex geometry and light conditions, and show good results using the proposed Daubechies basis on both synthetic and real datasets.