Object Detection on NCS ?

[idata]

Hi,

Just wanted to know if any of you have tried to get the object detection frameworks such as faster-RCNN or YOLO on the Neural compute stick.

If yes could you please share your experience/point to any references on the FPS you were able to achieve using VGG-16 net for feature extraction.

thanks,

jsb

[idata]

I was trying to use mvNCCompile to convert the caffemodel and prototxt of faster RCNN network by changing the PYTHONPATH to point to the "/py-faster-rcnn/caffe-fast-rcnn/python" branch included the faster-RCNN codebase(https://github.com/rbgirshick/py-faster-rcnn).

I had to do this as the region pooling layer required for object detection in the faster-RCNN framework is not included in the main caffe repo and is defined only in the caffe branch in the faster-RCNN repo. this caffe codebase is compiled and runs correctly in laptop.But when I try to run the following :python3 $NCS_TOOLKIT_ROOT/mvNCCompile.pyc /py-faster-rcnn/models/pascal_voc/VGG16/faster_rcnn_end2end/test.prototxt -w /py-faster-rcnn/data/faster_rcnn_models/VGG16_faster_rcnn_final.caffemodel -o ../networks/fasterRCNN/graph -s 12

I get the error :

Caffe Import Error

Could you please suggest what additional steps need to be done to point to a modified CAFFE installation in non-default location ? This would be useful if we wish to use any custom types of layers that are not yet included in the main branch of caffe yet.

Thanks,

jsb

[idata]

@jsb07 Currently we are looking into enabling object detection so stay tuned about this. All available layers in NCS are described in SDK Release Notes document but defining custom layers is not supported. About using caffe in different location maybe this discussion will be helpful to you

https://ncsforum.movidius.com/discussion/183/lapack-function-dgetri-lwork-could-not-be-found

[idata]

any update on this, i.e. support for faster rcnn on NCS?

[idata]

@manojgeorge Thanks for expressing your interests. Support for RCNNs is definitely something we are looking into, but there currently isn't a roadmap for this feature at the moment.

[ATell6]

Hello all;

I have a fasterRCNN and a fastRCNN gesture detectors implemented in MATLAB. I have only been successful exporting the fastRCNN network to ONNX.

I would like to use this ONNX model in NCS2.

I only see implementations of faster RCNN detectors as examples.

Could anyone provide an example of Fast RCNN detector for NCS2, where I can use my ONNX model after importing it to the format expected by the Intel device?

Thank you all in advance,

Alberto