Point Based Rendering of Multi-View Stereo Depth Maps

The recent progress in computer vision and multi-view stereo techniques permit to extract geometry from ordinary photo collections obtained from internet photo-sharing websites. This makes it easier than ever to capture highly detailed geometry of physical objects. In this work, different point-based methods are investigated to render and process the depth maps resulting from such a geometry acquisition technique. We present a multiresolution rendering approach well-suited for a larger number of depth maps, taking into account the redundancy and greatly varying sampling rate of depth maps. The method decomposes the input geometry according to the resolution of samples and stores them in multiresolution data structure. This data structure is used for memory efficient storage and level of detail selection throughout rendering. In a second, quite different approach, we adapt sequential point trees to render depth maps, thereby utilizing the graphics processor also for level of detail selection. Furthermore, we also investigate the resampling of depth maps by castings rays and intersecting them with a moving least squares surface