A sliced synchronous iteration architecture for real-time global stereo matching

In this paper, we present a low memory-cost message iteration architecture for a fast belief propagation(BP) algorithm. To meet the real-time goal, our architecture basically follows multi-scale BP method and truncated linear smoothness cost model. We observe that the message iteration process in BP requires a huge intermediate buffer to store four directional messages of the whole node. Therefore, instead of updating all the node messages in each iteration sequence, we propose that individual node could be completed iteration process in ahead and consecutively execute it node by node. The key ideas in this paper focus on both maximizing architecture's parallelism and minimizing implementation cost overhead. Therefore, we first apply a pipelined architecture to each iteration stage that is executed independently. Note that pipelining makes it faster message throughput at a single iteration cycle rather than consuming whole iteration cycle time as previously. We also make multiple message update nodes as a minimal processing unit to maximize the parallelism. For the multi-scale BP method, the proposed parallel architecture does not cause additional execution time for processing the nodes in the down-scaled Markov Random Field(MRF). Considering VGA image size, 4 iterations per each scale and 64 disparity levels, our approach can reduce memory complexity by 99.7% and make it 340 times faster than the general multi-scale BP architecture.