DNN module efficiency is low in C++

asked 2018-02-14 03:39:47 -0600

zrbzrb1106 gravatar image

I'm trying to run "master/samples/dnn/ssd_mobilenet_object_detection.cpp" of caffe module, and the efficiency is not high. I have tried to recognize a local video file but the FPS is only around 4-5. The Inference time is about 150ms. In the link below I found that the efficiency of DNN, C++ is good.

link text MobileNet-SSD @ 300x300 20 classes, Caffe 22.71 54.36 27.79

My CPU is I7-7700K 4.2GHz with 8 cores. Could someone give me some help? Thanks!

edit retag flag offensive close merge delete

Comments

@zrbzrb1106 please add a build configuration and how you estimate an efficiency. Have you tried to run OpenCV performance tests by ./opencv_perf_dnn --gtest_filter=DNNTestNetwork.MobileNet_SSD_Caffe?

dkurt gravatar imagedkurt ( 2018-02-15 02:08:54 -0600 )edit

Thank you for the reply. Actually I'm using the pre-built libraries in Windows10 and VS2017. The estimation is based on the output when running ssd_mobilenet_object_detection.cpp. I could see the output info of inference time and fps. I'm quite new in OpenCV, do I need to build the library myself to get higher efficiency? Thanks!

zrbzrb1106 gravatar imagezrbzrb1106 ( 2018-02-15 06:30:46 -0600 )edit

I have the same problem but when I'm trying to run resnet_ssd_face.cpp of caffe module in this case, the FPS is only around 1-1.8

system: my CPU is i5-3230M 2.60Ghz / windows8.1/ VS2017 And I'm using pre-built lib.

is this problem because of the pre-built lib or my system?

soheil gravatar imagesoheil ( 2018-12-22 18:45:30 -0600 )edit

@soheil , please do not post answers here, if you have a question or comment, thank you.

berak gravatar imageberak ( 2018-12-23 06:16:20 -0600 )edit
add a comment