Memory-efficient belief propagation for high-definition real-time stereo matching systems

Abstract

Tele-presence systems will enable participants to feel like they are physically together. In order to improve this feeling, these systems are starting to include depth estimation capabilities. A typical requirement for these systems includes high definition, good quality results and low latency. Benchmarks demonstrate that stereo-matching algorithms using Belief Propagation (BP) produce the best results. The execution time of the BP algorithm in a CPU cannot satisfy real-time requirements with high-definition images. GPU-based implementations of BP algorithms are only able to work in real-time with small-medium size images because the traffic with memory limits their applicability. The inherent parallelism of the BP algorithm makes FPGA-based solutions a good choice. However, even though the memory traffic of a commercial FPGA-based ASIC-prototyping board is high, it is still not enough to comply with realtime, high definition and good immersive feeling requirements. The work presented estimates depth maps in less than 40 milliseconds for high-definition images at 30fps with 80 disparity levels. The proposed double BP topology and the new data-cost estimation improve the overall classical BP performance while they reduce the memory traffic by about 21%. Moreover, the adaptive message compression method and message distribution in memory reduce the number of memory accesses by more than 70% with an almost negligible loss of performance. The total memory traffic reduction is about 90%, demonstrating sufficient quality to be classified within the first 40 positions in the Middlebury ranking.