Analysis of adaptive support-weight based stereo matching for hardware realization

Adaptive support-weight algorithm can generate high quality disparity map for stereo matching. But due to the complexity, it requires large internal memory size and bandwidth to meet the real-time constraint. In this paper, we first analyze the requirements of this algorithm from the hardware perspective. Then we propose our Support-Weight Window Reuse (SWWR) technique which can shorten computation time by the number of disparities, and Left-Right Cost Reuse (LRCR) to achieve bandwidth reduction by more than half. The comparison states that our proposed flow can generate much better disparity results, and meets the real-time constraint with relatively low memory cost and bandwidth.