A GPU Accelerator for Domain Transformation-Based stereo Matching
Qiong Chang, Aolong Zha, Masaki Onishi, Tsutomu Maruyama · 2019
Stereo Matching, which is a key technique for the depth detection in computer vision, can be widely used in many applications such as SLAM and Auto-driving. High accuracy, high processing speed and mobility of the stereo matching system are required in these ap- plications. Recently, with the development of artificial intelligence technology, a large number of stereo matching algorithms have been proposed, and the accuracy has been constantly improved. However, due to their high computation complexity, most of the current systems require to be run on large-scale hardware platforms. Hence, some researchers join the study of accelerating the stereo matching on the mobile platforms. In this paper, we propose a novel method to accelerate the state-of-the-art stereo matching algorithm Domain Transformation on a small and low power consumption GPU. Our method mainly focuses on improving the efficiency of data transfer between the multi-hierarchy memories of GPU during the cost aggregation step, and it consists of three optimization steps: Weight Generation, Aggregation Cost Penalizing and Wavelength Encoding & Decoding, which can help us solve the bandwidth waste caused by using floating-point data type due to the cost accumulation in Domain Transformation. By using this method, our system achieved 35 fps on Jetson TX2 GPU for 1280x375 pixels images when the maximum disparity is 128.