search

openvinotoolkit / open_model_zoo

Pre-trained Deep Learning models and demos (high quality and extremely fast)
https://docs.openvino.ai/latest/model_zoo.html
Apache License 2.0
4.1k stars 1.37k forks source link

How do we know the IR models details available at OpenVINO model zoo #756

Closed neso613 closed 4 years ago

neso613 commented 4 years ago

Hi

I am confused how would to know about the details and descriptions of IR files available at OpenVINO models Zoo?

Please suggest some way.

vladimir-dudnik commented 4 years ago

Hi @nehasoni3

what kind of details you are looking for? Ususally, all models in OMZ have some brief description, like for example this one. Is that what you are looking for?

Regards, Vladimir

neso613 commented 4 years ago

Hi @nehasoni3

what kind of details you are looking for? Ususally, all models in OMZ have some brief description, like for example this one. Is that what you are looking for?

Regards, Vladimir

HI @vladimir-dudnik OMZ folder have some demo but there are varity of IR files available at OpenVINO zoo. For example, there are 4 face detection IR files available, what are difference among them, what are backbone architecture, trained on TF/PyTourch/Caffe etc.

snosov1 commented 4 years ago

For the specific question about the difference of 4 face detection models

face-detection-adas-0001 is more accurate and face-detection-retail-0004 is faster.

face-detection-adas-binary-0001 is derived from face-detection-adas-0001 (i.e. similar accuracy) but some layers are binary which can give a slight performance increase on some platforms.

face-detection-retail-0005 is similar to face-detection-retail-0004 in terms of accuracy/performance, but it was trained in PyTorch (vs Caffe).

neso613 commented 4 years ago

For the specific question about the difference of 4 face detection models

face-detection-adas-0001 is more accurate and face-detection-retail-0004 is faster.

face-detection-adas-binary-0001 is derived from face-detection-adas-0001 (i.e. similar accuracy) but some layers are binary which can give a slight performance increase on some platforms.

face-detection-retail-0005 is similar to face-detection-retail-0004 in terms of accuracy/performance, but it was trained in PyTorch (vs Caffe).

Thanks @snosov1 How do you know all this. Please share.

snosov1 commented 4 years ago

We've developed the models, that's how =)

But as @vladimir-dudnik points out - every model has a description file. This information can be derived from there, I just gave it to you in a compact form.

neso613 commented 4 years ago

We've developed the models, that's how =)

But as @vladimir-dudnik points out - every model has a description file. This information can be derived from there, I just gave it to you in a compact form.

@snosov1 Okay!! You have developed these model. Its a request to please bring 68 points or more points facila landmark model on board. facial-landmarks-35-adas-0002 gives just two points for Eye. landmarks-regression-retail-0009 gives one point for Eye.

snosov1 commented 4 years ago

Sorry, @nehasoni3, this is not, currently, in our plans. The largest number of facial landmarks we have is 35 (facial-landmarks-35-adas-0002). I suggest to look for other open-source alternatives if you need more landmarks.

vladimir-dudnik commented 4 years ago

@nehasoni3, I would like to add, that you may contribute to OMZ an appropriate open source model, just follow contributing guide. We would be happy to review such pull request.