Any news on support for custom object detectors?

[idata]

Hi guys,

New to the community I am absolutely baffled about the performance on the RPi3+NCS. Especially for real time object detection.

I am looking in to using this on few coming projects instead of the Jetson TX2, but from what I can read it lacks support for e.g. models trained from TF Object Detection API.

Could you share with us the status of this? Being able to deploy custom trained TF detection models directly on the NCS would indeed be a game changer!

Thanks,

Jesper

[idata]

@JChr89 We don't have examples for object detectors for TensorFlow, but we do have sample code for TinyYolo and SSD Mobilenet Caffe @ https://github.com/movidius/ncappzoo.

[idata]

@Tome_at_Intel thanks for your response.

As far as I understood these examples are pretrained nets, and don't support custom architectures, e.g. transfer learning towards other objects. Is this not correct?

[idata]

@JChr89 The examples included in the appzoo and in the SDK are pretrained nets, but that does not mean you can't use transfer learning with them to classify/detect other objects. It is entirely possible to use transfer learning with the pretrained networks. In fact, AshwinVijayakumar has a post about transfer learning with GoogLeNet @ https://movidius.github.io/blog/deploying-custom-caffe-models/.

[idata]

@JChr89 Hey. Did you manage to find a solution to this? I have been on the lookout for an answer for this, but I'm struggling to find anything that's really useful to me. Thanks