Darknet (YOLOv3) & OpenVINO

[G__hany]

Hey all,

I successfully trained a model with 2 classes on darknet net (Tiny YOLOv3) and converted it to OpenVINO, and I am able to run it on their examples and even the OpenVino-YOLOv3 github repository.

2 main issues I've seen:

1. Convering darknet .weights -> .pb -> IR results in a very bad loss of accuracy (FP16), trying to fix the LeakyReLu (ie: 0.1000000023432 to 0.1 to 1.1) didn't really give any change. (the darknet model works better  compared - higher accracies) any idea what to do?

2. on darknet model, the inference indeed give me 2 classes identification (object detect) but when converting to IR, (with the .cfg) it can detect only 1 class! why?

Hope someone can help 

[yu__rui]

Hi, have u solved your first issue, my yolo model(FP16)'s acc drops from 95+ to 80+!