target tracking model

[yu__jin]

Hello

I recently installed the OpenVINO 2019R3 and all the dependencies on my laptop running ubuntu18.04. I built the demo successfully and run well for MYRIAD target case.

Now, I want to use OpenVINO for a new application about target tracking, not person-reidentification. 
I have my pytorch model based siamese network like siamfc、siamrpn、siamrpn++.

I was wondering if  the OpenVINO Toolkit support these networks.
and if there is any tutorial (beyond that 30+ demos ) to teach me how to implement OpenVINO for new applications of tracking? tracking not by redetection.

Thanks,

[Sahira_Intel]

Hi Jin,

OpenVINO should support your Siamese network, although there aren't any samples I can point you towards for reference. So long as your model does not have unsupported layers or operations, it should work. Here is the documentation for converting an ONNX model. 

In terms of tracking samples, can you please be more specific as to what kind of tracking you would like to achieve? This might be a good place to start.

Please let me know if this information is helpful.

Best Regards,

Sahira 

[yu__jin]

Hi Sahira:

I got a tracking model file( .ckpt) of tensorflow.   

And i run the command:

sudo python3 mo.py --input_meta_graph /home/btoob/disk/SiamFC-TensorFlow-master/scripts/Logs/SiamFC/track_model_checkpoints/SiamFC-3s-color-scratch/model.ckpt-0.meta --input_shape [1,255,255,3]

And i got these errors:

[ ERROR ]  Original placeholders: 'inference/examplar_input, inference/instance_input'. Freezing was requested for ''. --input_shape was provided without --input. Can not deduce which node shape to override
Exception occurred during running replacer "REPLACEMENT_ID" (<class 'extensions.front.user_data_repack.UserDataRepack'>): Original placeholders: 'inference/examplar_input, inference/instance_input'. Freezing was requested for ''. --input_shape was provided without --input.

Any help would be greatly appreciated

[Hu__Shuping]

Hi Sahira,

I'm trying to infer siamrpn/siamrpn++ network in openvino(.xml and .bin was converted from .onnx successfully), and error "layer n post input 1 is not connected to any data " occurs in openvino version 2020.1, "Incorrect number of input edge for layer n" in openvino version 2019, it seems that there has no unsupported layers in the network. This network has 2 inputs (denote as z and x). The only special issue in layer n is that the kernel weights in this layer is not allocated in the bin file but allocated dynamically (get the output feature map in layer n-1 of input z as the kernel weights of layer n, get the output feature map in layer n-1 of input x is the input of layer n). Does openvino support this kind of operation? 

Sahira R. (Intel) wrote:

Hi Jin,

OpenVINO should support your Siamese network, although there aren't any samples I can point you towards for reference. So long as your model does not have unsupported layers or operations, it should work. Here is the documentation for converting an ONNX model. 

In terms of tracking samples, can you please be more specific as to what kind of tracking you would like to achieve? This might be a good place to start.

Please let me know if this information is helpful.

Best Regards,

Sahira