Intel® Distribution of OpenVINO™ Toolkit
Community support and discussions about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all things computer vision-related on Intel® platforms.
5924 Discussions

Different inference result between Movidius 1 and Movidius 2

junyi
Beginner
224 Views

OS Platform: Ubuntu 16.04.5 LTS

OpenVINO Version: 2018 R5 (Build date:17 Dec 2018)

 

When running classification_sample, I got unexpected result when using different hardware

I using following command to run classification_sample on CPU

./classification_sample -d CPU -m model.xml -i image.png

and following command on Movidius 1 and Movidius 2

./classification_sample -d MYRIAD -m model.xml -i image.png

I was satisfied that OpenVINO could give me consistent result when using different hardware,

but when I test with Movidius 2 I got unexpected result that didn't align to CPU & Movidius 1

 

CPU Top 5 results

Image image.png

4 1.0000000 label #4

1 0.0000000 label #1

8 0.0000000 label #8

9 0.0000000 label #9

7 0.0000000 label #7


Movidius 1 Top 5 results

Image image.png

4 1.0000000 label #4

1 0.0000000 label #1

8 0.0000000 label #8

9 0.0000000 label #9

7 0.0000000 label #7


Movidius 2 Top 5 results

Image image.png

3 0.3332520 label #3

8 0.3332520 label #8

9 0.3332520 label #9

7 0.0000000 label #7

1 0.0000000 label #1

My question is what cause the inconsistent result

0 Kudos
3 Replies
anchor__jiang
Beginner
224 Views

I met the save problem! 

Tikhostoup__Dmitri
224 Views

We have similar problem, certainly some NN models compute different, where NCS v2 delivers incorrect results. 

My colleague posted detailed description of the problem here: https://software.intel.com/en-us/forums/computer-vision/topic/801760

almost a month ago, still no reply from Intel.

 

Jakub
Beginner
224 Views

I also faced the problem of incorrect predictions using NCS2 (described at https://software.intel.com/en-us/forums/computer-vision/topic/802216). In my case, I found a workaround of manually editing the XML file generated by the Model Optimizer, but it only works if the model is placed on CPU, not MYRIAD.

Reply