Can the PRL Tracker model be deployed on the Metis Accelerator?

Hi ​@Eva , great to see you here!

We’ve chatted a little about this in the background, but I thought it might be helpful to follow up here too.

So, the PRL Tracker isn’t part of our Model Zoo, and I've not seen anyone on the community talking about deployment of the PRL Tracker model on Metis yet.

However, our internal tools and SDK do support similar PyTorch-based models. I'd say it's technically feasible yep, but it may well require a fair amount of adaptation effort, especially around model conversion and pipeline integration.

But maybe ​@Victor Labian, ​@Habib or ​@Jonas might have a more informed idea about its feasibility?

hi ​@Eva 
We support some of the mentioned backbone like resnet and mobilenetv2. 
As first step you can try to compile the models you want to use. 
https://github.com/axelera-ai-hub/voyager-sdk/blob/181add2c01593550f6d98f8ec48fa2e9d1acbe7b/docs/reference/compiler_api.md#python-compiler-api
 

If these work, you can build your own pipeline using axruntime or axinferencenet 

https://github.com/axelera-ai-hub/voyager-sdk/blob/181add2c01593550f6d98f8ec48fa2e9d1acbe7b/docs/reference/axelera.runtime.md
https://github.com/axelera-ai-hub/voyager-sdk/blob/181add2c01593550f6d98f8ec48fa2e9d1acbe7b/docs/reference/axinferencenet.md

​@Eva This is very interesting. I have been myself looking for a visual tracking algorithm that can be run on edge devices with good performance. Have you been able to run this on any edge platform so far? I am curious what kind of performance you get from the model in terms of fps.