Hello Im trying to use my own trained model of yolov3-tiny in OpenCV. Im have code from this tutorial , but when im using videos or camera then this is going very slow(300ms) but when i`m using this same files cfg and weights in darknet then i have 40-50 fps. Is there any way to speed up this c++ code to get this same result, except lowering resolution in code, becouse it has normal speed only for 32x32 but then it detect nothing.
Like stated by @gino0717, OpenCV DNN backend is not yet optimized with CUDA, which means it runs on CPU if you did not enable the OpenCL backend. However, there are optimized CPU models using the OpenVINO backend, have a look here: https://software.intel.com/en-us/open...
yep that is possible, since they are acquired by Intel and probably Intel is pushing on that side. For the GSoC2019 program, there are students applying for integrating CUDA backends, so it might be slowly arriving ;)
Asked: 2019-04-14 12:25:36 -0600
Seen: 2,331 times
Last updated: Apr 15 '19
Do you have the same results for original yolov3-tiny from https://pjreddie.com/darknet/yolo/ ? I am experimenting with these models and object_detection.cpp example. Inference time is around 30ms for yolov3-tiny.
I guess you used GPU in darknet but CPU in openCV dnn
i needed change number of classes and filters and now this is working this same as mine