Tiny YOLO on NCS

[idata]

I've been trying to find a way to run an object detection network on the NCS. After a while trying and converting different models I finally found a one that seems to work; Tiny YOLO. I haven't yet found a way to convert the weights from darknet to caffe to NCS, but with empty weights the performance looks promising. Under 150ms per image!

python3 mvNCProfile.pyc yolo_tiny_deploy.prototxt -s 12

mvNCProfile v02.00, Copyright @ Movidius Ltd 2016

__ WARNING: using empty weights __

USB: Transferring Data…Time to Execute : 140.52 msUSB: Myriad Execution FinishedTime to Execute : 121.39 msUSB: Myriad Execution FinishedUSB: Myriad Connection Closing.USB: Myriad Connection Closed.Network Summary

Detailed Per Layer Profile

Layer Name MFLOPs Bandwidth MB/s time(ms)

0 scale1 173.408 1173.26 8.81

1 pool1 3.211 811.31 7.552 scale2 462.422 964.80 14.303 pool2 1.606 956.24 3.204 scale3 462.422 698.33 9.925 pool3 0.803 1008.54 1.526 scale4 462.422 428.05 8.387 pool4 0.401 1001.39 0.768 scale5 462.422 205.36 11.139 pool5 0.201 980.96 0.3910 scale6 462.422 308.66 10.0811 pool6 0.100 944.83 0.2012 scale7 462.422 809.99 11.6413 scale8 231.211 596.90 8.9814 fc9 0.025 2146.64 16.40 Total inference time 113.27

Does anyone have experience converting the weights? Or knowledge how to draw the bounding boxes based on the YOLO output layer? It appears to output 1470 values.

[idata]

@victorv, yes, I mean full yolov2. The layer of reroute and reorg should be supported by NCS at least.

[idata]

@xhuan28, agree with @victorv. I don't plan to release anything publicly, but you can make any object detector model work by executing custom layers within Caffe. Usually, custom layers are not that heavy to process. Ideally, Movidius should integrate calls to NCS into Caffe using e.g. engine: NCS parameter.

[idata]

MDK licensees could write custom Caffe layers for NCS - not sure if non-MDK users can do this.

Reorg layer in darknet is the same as Reshape in Caffe. The route layer in darknet is not supported in current released Caffe. You can safely drop the route layer with minimal performance impact on time and mAP.

[idata]

This is a brilliant project. I basically bought this thing to run Yolo on it and here you've done it :) Thanks so much @dagofff .

[idata]

I have tried to generate custom model of yolo tiny for ncs, but cannot make the model that can convert to caffemodel by the prototxt. Which cfg file to train on darknet?

[idata]

I create a repo for running yolov2 for ncs in https://github.com/duangenquan/YoloV2NCS. It contains some c/c++ implementation for the cregionlayer, a pythonwrapper and model convertor for yolov2 to caffe. Hope this helps.

[idata]

that is good for me also.

could you share movie clip about runnig yolo with ncs?