Make use of the .pb file from the pretrained model

Make use of the .pb file from the pretrained model

Thanks for such an amazing tool for pose estimation.

I’m building an action classifier which runs on realtime USB camera and need to use OpenVINO toolkits to deploy the model on VPU to speed things up.

(If you’ve heard of Neural Compute Stick 2 by Intel, VPU is basically the same thing. You can deploy the inference task on VPU to lower the comsumption of CPU and speed up the inference.)
However this means I ONLY need the pb file from the model I’ve trained and need not further assistance from the DLC api like video analysis.
I need to get the position of each keypoint detected and feed these into another classification network for further analysis.
However, I looked into the model with Tensorboard and I couldn’t really understand the network structure, which is shown as below.
BTW I’m using the pre-trained full-dog model from the deep model zoo.

It seems like cutting the graph from the “pose” node (assign this node as output) what I needed but I’m not really sure.

Can anyone tell me what the purpose of operations done after this “pose” node is? Or anyone interested in the usage of OpenVINO with me, feel free to contact me by my email addresss or by commenting below.


You can simply run export_model and get the protobuffer file. See our new paper here: