I own two Movidius sticks, a NCS1 and a NCS2. Both seem to work fine (the example applications show correct results). However, when I convert my own Tensorflow models using the model optimizer, I get wrong results using the NCS2 while getting correct results on the NCS1.
What could be the reason for this?
To reproduce this issue, I created a small github repo containing the converted model as well as all code and commands used to train, freeze and convert the model: email@example.com:kelia/movidius_tensorflow.git
If you want to reproduce the issue, do the following steps:
1. Clone repo
2. Install dependencies (openvino, opencv, tensorflow, colored)
"sudo pip3 install tensorflow opencv-python colored"
3. Plug in Movidius stick
4. Run "python3 compare_results.py"
The script prints the results of 4 inference variants: plain tensorflow, froze graph, openvino myriad and openvino cpu. All 4 results should match, but on my machine the result for myriad on the NCS2 is wrong.
Thanks a lot for any help!