Key points are not available for this paper at this time.
In future 3D videoconferencing systems, depth estimation is required to support autostereoscopic displays and even more important, to provide eye contact. Real-time 3D video processing is currently possible, but within some limits. Since traditional CPU centred sub-pixel disparity estimation is computationally expensive, the depth resolution of fast stereo approaches is directly linked to pixel quantization and the selected stereo baseline. In this work we present a novel, highly parallelizable algorithm that is capable of dealing with arbitrary depth resolutions while avoiding texture interpolation related runtime penalties by application of GPU centred design. The cornerstone of our patch sweeping approach is the fusion of space sweeping and patch based 3D estimation techniques. Especially for narrow baseline multi-camera configurations, as commonly used for 3D videoconferencing systems (e.g. 1), it preserves the strengths of both techniques and avoid their shortcomings at the same time. Moreover, we provide a sophisticated parameterization and quantization scheme that establishes a very good scalability of our algorithm in terms of computation time and depth estimation quality. Furthermore, we present an optimized CUDA implementation for a multi GPU setup in a cluster environment. For each GPU, it performs three pair wise high quality depth estimations for a trifocal narrow baseline camera configuration on a 256x256 image block within real-time.
Waizenegger et al. (Thu,) studied this question.