Eric Viscito, NVIDIA
GTC 2020
Computer vision algorithms in many applications, such as video analytics, robotics, and autonomous vehicles, depend on motion and depth information. Recent GPU SOCs include a new hardware engine—the Optical Flow Accelerator (OFA)—that accelerates the estimation of optical flow and stereo disparity, freeing up the GPU from these tasks. The engine is exposed through high-level APIs and can be integrated into various computer-vision pipelines. We'll introduce the OFA engine, teach its principles and modes of operation, discuss throughput and quality metrics, and illustrate some applications. We're aiming for the general computer-vision engineer, as well as any GPU software engineer desiring a fuller understanding of the capabilities of modern GPUs. In-depth knowledge of optical flow or stereo disparity estimation algorithms is not required, although some pre-existing knowledge in those areas will enable a better appreciation of the principles of operation.