We are using Yolo Model (v3, Tiny yolo) for object detection and tracking. This model gave us extremely good accuracy when we ran it on GPU, however to deploy the same on NCS2 (purchased recently) we converted the yolo model to pb file then to IR file and ran it on NCS2, Now the accuracy is extremely bad in terms of object detection.
Kindly let us know so as to how to avoid the loss of converting the yolo model to pb file and then to IR (FP 32)
Thanks and regards
Dr. Pallavi N Halarnkar
Thanks for reaching out! The Intel Neural Compute Stick 2 only supports FP16, could you please answer the following questions?
- Which version of the OpenVINO toolkit are you using?
- How did you train your model? Darknet?
- What command did you use with the model optimizer?
- Would it be possible to share your model to test from my end?
Let me know if you would like to share the files privately and I can send you a private message.