New option to convert YoloV3 .weights file to OpenVINO IR model

[Tsin__Ross]

See https://github.com/rossqin/RQNet

Convert to RQNet model and then IR Model

1. RQNet wconv -c <path/to/darknet/network/config> -i <path/to/darknet/weights> [-o <dir/for/output>]

2. RQNet openvino -n <path/to/rqnet/network/defintion> -w <path/to/rqnet/weights> [-o <dir/for/output>] [-d FP16|FP32] [-t model_name]

I tried. It works perfectly.

Let me know if you have questions or concerns.

Enjoy~~~

[Shubha_R_Intel]

Dear Tsin, Ross,

Thank you for sharing such gold nuggets with the community ! 

We appreciate it,

Shubha

[Tsin__Ross]

Thank you Shubha.

I used Tiny YoloV3, but as far as I see, the model converted and run on NSC2 does not reproduce the precision as on my GPU.

Here I posted my  codes, could you please help me confirm this? I doubt there may be some kind of layer which only used by Yolo has defects(since the MobileNet +SSD model works), but not sure.

I am now developing a CV based drone, and NCS2 is very useful in the project(in my verification codes) I get around 10fps on Raspberry Pi+NCS.

Hope I can do something to contribute to the eco-system of OpenVINO.

Cheers,

Ross

Shubha R. (Intel) wrote:

Dear Tsin, Ross,

Thank you for sharing such gold nuggets with the community ! 

We appreciate it,

Shubha