Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.
6392 Discussions

Has anyone successfully converted/used the TensorFlow Object Detection models on Movidius?

idata
Employee
631 Views

Hi all

 

_[disclaimer: noob to ML in general]_

 

I'm getting started with the NCS and can successfully run demos using an SSD model for caffe generously supplied by github:chuanqi305

 

https://github.com/chuanqi305/MobileNet-SSD

 

But AFAICS that model was trained with 21 classes, while the one provided by the TensorFlow team includes 99 classes: some of the additional classes are ones I need for my application (for example: cup/glass)

 

'http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_11_06_2017.tar.gz

 

So I thought that a reasonable next step would be to compile that TF model, which I've used with my local TF, to be usable on NCS.

 

But I also thought that it's likely that someone else has done this or has compiled a model that has a larger number of classes.

0 Kudos
3 Replies
idata
Employee
419 Views

@raygeeknyc At the moment, the NCSDK doesn't have support for SSD MobileNet on TensorFlow. The main difference that you're mentioning is the dataset that the network was trained with. Likely the 21 class SSD MobileNet was trained with the VOC dataset and the 99 classes version was trained with the COCO dataset.

0 Kudos
idata
Employee
419 Views

Thanks! Ok my question now becomes, does anyone here have a model trained with the COCO dataset?

0 Kudos
idata
Employee
419 Views

@raygeeknyc You can try to convert tensorflow model to caffe model first~

0 Kudos
Reply