0

Hi I'm trying to create landmarks from the blazeface model with openvino. The model is running and I have in regression output a shape array of (1,896,16) with these values corresponding to tensors:


[10.116998 2.7768276 29.180841 29.190268 -7.6588407 -1.8769206 7.955852 2.8945901 -1.1223608 11.560151 4.533435 13.420406 -7.406376 -9.7577505 25.858068 -1.905251 ]

[10.240653 2.762262 29.685017 29.694557 -7.6490545 -2.103835 8.018975 2.6713266 -1.0454882 11.4031725 4.6832647 13.421963 -7.434411 -9.877215 26.122345 -2.0081995]

[10.195236 2.7623253 29.391754 29.400812 -7.406289 -2.2815979 7.871913 2.3147857 -1.0604903 10.664557 4.600431 12.801059 -7.033283 -9.625556 25.726923 -1.9212234]

[ 9.646918 2.7344685 28.515514 28.524525 -7.4333186 -1.8820286 7.6982303 2.9067798 -1.1013308 11.309927 4.33471 13.137727 -7.308227 -9.689351 24.914877 -1.7770973]

The problem is I don't know what kind of value the blazeface model output give me and how to create landmarks from that. I saw they use get_tensor function after regression results in tensorflow but I didn't find the equivalent with openvino:

https://gist.github.com/ogl4jo3/5ba673b61ecc82780299edbcdcca301a

If someone has an idea please.

1 Answers1

0

I had a brief look at the sample you provided and saw that get_tensor is used to get raw results from network. To get boxes you need to use ProcessCPU and orig_nms

I would suggest you to use Python API of Inference Engine from OpenVINO. An example of usage is the link

You can try to adapt the sample above for your use case and use the functions ProcessCPU and orig_nms to extract boxes from raw data.

Artemy Skrebkov
  • 346
  • 1
  • 8