High Quality and Memory Efficient Representation for Image Based 3D Reconstructions

Tobias Nöll, Johannes Köhler, Gerd Reis, Didier Stricker

Abstract: We propose a practicable, fully automatic pipeline that directly computes high quality and memory efficient geometry representations from image based 3D reconstructions. It takes a set of calibrated depth and texture images as input. First, we iteratively compute a low frequency geometry proxy directly from the depth images in a photoconsistent way. Second, the large amount of high frequency detail information from all input images is consistently aggregated in a texture and a normal map. View dependent illumination artifacts such as specular reflections are removed during this process. The separation of the underlying massive amount of data into a general shape proxy and high frequency information maps yields a both efficient and visually pleasant representation. Our output can directly be used by digital artists or integrated in memory critical realtime applications such as computer games or web based visualizations.