简体繁体 English

OpenVX光流金字塔性能

[英]OpenVX optical flow pyramids performance

原文 2015-04-10 17:56:22 9 2 opencv/ computer-vision/ opticalflow/ openvx

The question is about performance of OpenVX compared to the traditional OpenCV package. 问题是与传统的OpenCV软件包相比，OpenVX的性能如何。 Do you know guys, which of them is faster in the calculation of optical flow pyramids? 您知道吗，在光流金字塔计算中，哪个更快？ The function in OpenCV for this operation is following: cv::buildOpticalFlowPyramid(nextFrame, nextPyr, winSize, maxLevel, true); OpenCV中用于此操作的函数如下：cv :: buildOpticalFlowPyramid（nextFrame，nextPyr，winSize，maxLevel，true）; Who is faster? 谁更快？

2 个解决方案

OpenVX is just a standard. OpenVX只是一个标准。 You can compare performance of some OpenVX implementation. 您可以比较某些OpenVX实现的性能。 As far as I know, currently only sample implementation of the OpenVX standard (can be found on official Khronos site) is available publicly. 据我所知，目前仅公开提供OpenVX标准的示例实现（可在Khronos官方网站上找到）。 This sample implementation doesn't use any optimizations, so I suppose it will be slower then OpenCV. 此示例实现未使用任何优化，因此我想它会比OpenCV慢。 To get real comparison you need to wait official production implementation of OpenVX standard from one of vendors. 为了进行真正的比较，您需要等待其中一家供应商提供的正式生产OpenVX标准。

I tested openVX feature detection (harris corner detection) and tracking (LK optical flow Pyramid) on Jetson TX2 GPU and got algorithm runtime in the range of 2-5ms (based on the features in the video, in my case moving car was capturing features on the road). 我在Jetson TX2 GPU上测试了openVX特征检测（哈里斯角点检测）和跟踪（LK光流金字塔），并且算法运行时间在2-5ms范围内（基于视频中的特征，在我的情况下，移动汽车正在捕获特征）在路上）。

I am not sure if OpenCV CUDA version can beat this. 我不确定OpenCV CUDA版本是否可以胜过这个。