Movidius NCS2 low inference accuracy problem?

[robert_osmosis]

I have a dataset that I am training with Tensorflow's object detection model (SSD Mobilenet v1) and I am experiencing significant accuracy loss when converting for the movidius stick using the model_optimizer included with the OpenVINO SDK

The same dataset performs well when trained with the tiny yolov3 model in keras but with the NCS2 the boxes are giving out a lot of false positives and the boxes around the objects are too big.

Has anyone experienced something similar?

[Ansif_M_Intel]

Hi,

A small decrease in accuracy is expected, however if you are getting a marginal decrease in accuracy kindly revert back to us and share your workload. so that we can replicate the scenario at our end.

[Aswathy_C_Intel]

Hi, Could you please confirm if the issue is resolved. Please be informed that the thread will get closed within 2 business days assuming that the solution provided was helpful.

[Aswathy_C_Intel]

We are closing this thread. Please raise a new one if you face any further issues. After the case closure, you will receive a survey email. We appreciate if you could complete the survey regarding the support you received.