model performance on live feed

[pkhan10]

Hello I am using python api to call person detector using person-detection-retail-0013.xml

Code for person detection

def generate_person_detection(input_frame,exec_net,n=1,c=3,w=544,h=320,thresh=.1):
    print("number of request",len(exec_net.requests))
    op_frame = cv2.resize(input_frame,(w,h)).transpose((2, 0, 1)).reshape(n,c,h,w)
    ### we can add multiple requests and just enumerate request ids
    exec_net.start_async(request_id=1, inputs={input_blob: op_frame})
    if exec_net.requests[1].wait(-1)==0:
        res = exec_net.requests[1].outputs[out_blob]
    res_filt =  res[np.where(res[:,:,:,2]>thresh)]
    return res_filt

although in the begining fps is around 60 it decrease to very low value(10,1 ,0) over  the time refer the attachment or onedrive url Its a plot of fps vs frames.

I am testing it on core i7 , 7th gen processor

[Shubha_R_Intel]

Dear khandelwal, prateek,

Which version of OpenVino are you using ? If you're not using OpenVino 2019R2 please upgrade - it was just recently released. Performance improvements and enhancements are always made to each release. The best way to figure out what is wrong with your model is to experiment with the benchmark_app. Also are you using Async API ? (StartAsync).

You should be able to narrow down best performance using benchmark_app and playing with its different tuning parameters.  Also this Performance Topics Document will help.

Please report your progress on this forum and thanks for using OpenVino !

Shubha