[OpenVINO] Enable Native Image Segmentation C++ Demo

[mingmingtasd]

The source code of this Image Segmentation C++ Demo we have enabled blongs to openvino-toolkit, and it seems that the source code can't be found on github. While I find another Image Segmentation C++ Demo on https://github.com/opencv/open_model_zoo/tree/master/demos/segmentation_demo with similar source code and feature.

For the Image Segmentation C++ Demo we have enabled, excute this command ./segmentation_demo -i woman.bmp -m /home/xu/Desktop/deeplabv3_mnv2/without_argmax/deeplabv3_mnv2_224/deeplabv3_mnv2_224.xml, we can segment the first image to get the second image as output:

[fujunwei]

How to get the deeplabv3 openvino model? do you have any patches on the Image Segmentation C++ Demo ?

[mingmingtasd]

@fujunwei We converted tensorflow model with openvino tool to get this openvino model. The demo has been upgraded in openvino 2020.2 version and should to be rebuilt if needed.

[fujunwei]

Could you please share the patch you apply to the Demo? i want to verify it with old veriosn.

[mingmingtasd]

OK, I have sent to you privately. @fujunwei

[mingmingtasd]

On openvino 2020.3.194, the Native Image Segmentation C++ Demo has been updated. I tried to re-enable this demo to validate the deeplabv3 model (513*513) on ICL. I find that the updated demo code uses openCV to post-process output data but it doesn't do pre-processing for input. So add some codes to resize the input picture, for example:

cv::Mat res_inImg;
cv::Size sz;
sz.width = 513;
sz.height = 513;
cv::resize(inImg, res_inImg, sz);
......

In addition, must set -delay 0 in the command line as below: ./segmentation_demo -m /home/webnn/Desktop/quantize/deeplabv3_DefaultQuantization/2020-06-22_11-00-05/optimized/deeplabv3.xml -i ~/Downloads/woman.jpg -delay 0

The original deeplabv3 model and the quantized deeplabv3 model can both pass with the demo, check the details as below:

@ibelem @huningxin