Model Optimizer question

[tsai__ellen]

Hi,

Converting a TensorFlow* Model whether to support Faster R-CNN Vgg16 ?

Regards,

Ellen

[Shubha_R_Intel]

Dear tsai, ellen,

yes that is supported. Please study the supported tensorflow models and  also converting Tensorflow Object Detection API models .

Thanks,

Shubha

[tsai__ellen]

Dear Shubha

Thank you your reply : )

- Is All the Faster R-CNN 'xxxxx' supported?

I am still confused.
Is the 'Faster R-CNN Vgg-16' is equal to 'VGG-16' ?
In the Tensorflow supported list, the 'VGG-16' exists but the 'Faster R-CNN VGG-16' is not exist.
So, is nor matter what 'xxxx' is, is all the Faster R-CNN 'xxxxx' supported?

- the reason why I ask

I ask this question, because Faster R-CNN is supported in Caffe* framework, but not in Tensorflow* framework.
I want to make sure it support in Tensorflow* framework.
https://docs.openvinotoolkit.org/2018_R5/_samples_object_detection_demo_README.html (Caffe* Faster R-CNN VGG-16)

- situation 

At the moment, I can't successfully convert .ckpt to a .pb file. I plan to give up this model for a while, although its results are better.

Thanks,

Ellen

[Shubha_R_Intel]

Dear tsai, ellen,

I can certainly understand why you are confused. And your observation below is correct also:

In the Tensorflow supported list, the 'VGG-16' exists but the 'Faster R-CNN VGG-16' is not exist.

VGG-16 (and other VGG-x) is a CNN network architecture invented by The University of Oxford Visual Geometry Group .

This article explains it very well.

At the conceptual level, Faster-RCNN is composed of 3 neural networks — Feature Network, Region Proposal Network (RPN), Detection Network.

The Feature Network is usually a well known pre-trained image classification network such as VGG minus a few last/top layers. The function of this network is to generate good features from the images. 

So if you look at the names of the supported models on the Model Optimizer Tensorflow Supported List you will see stuff like : 

Faster R-CNN Inception ResNet V2 Low Proposals COCO which means that Inception ResNet V2 is used as the detection network but was trained with the COCO dataset

OR

Faster R-CNN Inception ResNet V2 Open Images* which means that Inception ResNet V2 is used as the detection network but was trained with the OpenImages* dataset,

and so on...

But keep in mind, Intel is not training or creating those Faster R-CNN models. The tensorflow community created the models on the supported list - however Model Optimizer happily converts them to IR.

So yes, we have VGG-16 and VGG-19 but these are not Faster R-CNN models. They are just plain old VGG models.

I hope my explanation makes sense.

Thanks,

Shubha