age-gender-recognition-retail-0013 through cv::dnn API [SOLVED]

[Bigorgne__Erwan]

Hello,

The embedded opencv release in openvino 2019 R2 provides a very convenient way to handle the pre-trained net. thanks a lot for that.

But for some reason, I can't get any consistent result with the age-gender-recognition-retail-0013 net (FP32 CPU, the same behavior is observed with FP16/MYRIAD)

                  theNet.setInput(blob,"data");
                  std::vector<cv::Mat> netoutputs;
                  std::vector<cv::String> layers{"age_conv3"};
                  theNet.forward(netoutputs,layers);
                  float age =netoutputs[0].at<float>(0)*100.;

where blob = cv::dnn::blobFromImage(...)

I can't figure out why. The age value takes  random values through the runs.

[Shubha_R_Intel]

Dear Bigorgne, Erwan, please study the code of the following demo.  This Demo shows Face Detection coupled with Age/Gender, Head-Pose, Emotion, and Facial Landmarks detectors. Supports video and camera inputs.

Interactive Face Detection C++ Demo

Hope it helps,

Thanks,

Shubha

[Bigorgne__Erwan]

Hi Shubha, thank you for the reply,

I eventually evaluated the age-gender net using the InferenceEngine API. It works. However, I would have liked to stick to opencv dnn API, as most of our functionnal blocks are based on this latter one.

Despite it seems to be an opencv-related problem, as openvino embeds a release of opencv libraries compiled by your team, I thought it was better to mention this problem here, in an Intel's forum.

All the openvino-zoo models but the age-gender one work great through the cv::dnn API...

Thanks again

Erwan

[Shubha_R_Intel]

Dear Bigorgne, Erwan,

This forum is dedicated to Model Optimizer and Inference Engine. If you are building your Deep Layer application using OpenCV, you will get better response in the OpenCV forum:

https://github.com/opencv/opencv/issues/

Thanks,

Shubha

[Kurtaev__Dmitry]

Hi!

Can reproduce it with a single output. However works fine if you get both age and gender:

image source: https://github.com/opencv/opencv_extra/blob/master/testdata/gpu/lbpcascade/er.png

[Dmitry_K_Intel3 · Accepted solution]

Hi!

Can reproduce it with a single output. However works fine if you get both age and gender.

image source: https://github.com/opencv/opencv_extra/blob/master/testdata/gpu/lbpcascade/er.png

[Bigorgne__Erwan]

Hi Dmitry,

You are right!! I may dig deeper to know why.

"If you can't get one one correct output, don't bother to get both of them" doesn't apply here...

Thank you very much.

Erwan.