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.
6058 Discussions

Object detection Model showing different output results on Myriad and CPU on same image inference

MohnishJain
Beginner
730 Views
I am trying to perform inference on a custom trained model for person detection. i am trying to perform inference on a image on CPU and Myriad Movidius.

The Detections result is different in both the cases.

Myriad Results:

results [[[-5.87500000e+00 -9.79687500e+00 3.38281250e+00 3.91796875e+00
1.38998032e-04 3.34960938e-01]
[-4.57812500e+00 -1.27421875e+01 3.24609375e+00 3.92968750e+00
0.00000000e+00 3.88427734e-01]
[ 3.67431641e-01 -1.29296875e+01 2.89648438e+00 3.69726562e+00
0.00000000e+00 4.71435547e-01]
...
[-4.85595703e-01 4.73437500e+00 -3.82812500e-01 6.41601562e-01
0.00000000e+00 9.86328125e-02]
[ 2.15039062e+00 6.93750000e+00 6.95312500e-01 1.01757812e+00
0.00000000e+00 1.17431641e-01]
[ 1.97167969e+00 8.57031250e+00 2.93164062e+00 3.54296875e+00
3.71932983e-04 2.06420898e-01]]]

CPU Results:
results [[[-8.4628897e+00 -1.3867578e+01 3.7331104e+00 4.3173060e+00
2.2710463e-04 4.3838483e-01]
[-5.8339972e+00 -1.9294828e+01 3.4993901e+00 4.2565918e+00
4.6186513e-05 5.2212775e-01]
[ 1.0963238e+00 -2.0203087e+01 3.1744423e+00 4.0058889e+00
2.3131161e-05 5.8755422e-01]
...
[-1.2445596e+00 7.6259720e-01 -4.1536269e+00 -2.3483934e+00
3.5924142e-05 8.0652714e-02]
[ 1.3845503e-01 2.4664059e+00 -2.0437424e+00 -1.5161291e+00
2.2524493e-05 1.0168492e-01]
[ 1.3065454e+00 6.8799105e+00 2.3910873e+00 2.9384508e+00
7.6559943e-04 1.8868887e-01]]]


Results Posted by CPU Inference are correct and movidius (myriad X NCS2 stick)results are wrong.


Kindly assist how to handle this scenario as results obtained on movidius are having higher standard deviation.


Download Path:

https://drive.google.com/drive/folders/1Wyw423tf7WE1YU_w7T0h3hdbusDdodAJ?usp=sharing

Command:
python demo\OpenVINO\python\demo.py

NOTE:
I have attached .pth ,onnx and IR models in it. you can try with both IR and ONNX models.
Labels (1)
0 Kudos
9 Replies
Hairul_Intel
Moderator
673 Views

Hi MohnishJain,

Thank you for reaching out to us.

 

I've replicated this issue and received similar results (I've tested with CPU and MYRIAD plugins on Windows 10).

 

We're investigating this issue and shall get back to you soon.

 

 

Regards,

Hairul

 


MohnishJain
Beginner
660 Views

@Hairul_Intel  Do You have any updates for this issue? we kind off needed the resolution/reason behind this issue on priority as we have to move this in production in couple of weeks.

 

So, we would appreciate if you can help a bit here.

Hairul_Intel
Moderator
617 Views

Hi MohnishJain,

Apologies for the delayed response.

 

Currently we are escalating this issue to the development team for further  troubleshooting. This should take a while, thanks for your patience so far.

 

 

Regards,

Hairul

MohnishJain
Beginner
509 Views

@Hairul_Intel  Any Update on this issue resolution, it has been 2 months  since the issue has been raised.I have been waiting long for this to get solved and temporary fixes are throwingissues now.

 

we are moving now in production, if the issue still exists we need to drop the movidius stick option for inferencing.

 

 

Hairul_Intel
Moderator
482 Views

Hi MohnishJain,

Apologies for the delay on this.

 

Our development team is still looking to rectify this and working to provide a robust solution. I will keep you updated once available.

 

Sorry for the inconvenience caused.

 

 

Regards,

Hairul

MohnishJain
Beginner
403 Views

@Hairul_Intel  Any Update on this it has been 4 months .since we have reported the issue

Hairul_Intel
Moderator
381 Views

Hi MohnishJain,

We sincerely apologize for the delay in providing updates for this issue.

 

Our development team is still working on resolving this issue. The fix will usually be embedded into the next upcoming release. We will keep you posted once we obtain the official timeline from the development team.

 

We appreciate your understanding in this matter and thank you for your continuous support of OpenVINO.

 

 

Regards,

Hairul

Hairul_Intel
Moderator
178 Views

Hi MohnishJain,

Glad to inform you that our development team has rectified the reported issue on our latest releases (OpenVINO 2022.1 onwards). We would like to invite you to test from your side.

 

Sharing our replication results using OpenVINO 2022.1 (with API 2.0) to compare inference results across CPU, GPU and Myriad as shown in the screenshot below.

 

2022 results.png

 

MO command:

mo --input_model public_model.onnx --data_type FP16 --layout NCHW --use_new_frontend --input_shape [1,3,1024,1024] --mean_values [0.485,0.456,0.406] --scale 255 --model_name demo_m_public --output_dir ir_models/2022.1/

 

 

Regards,

Hairul

 

Hairul_Intel
Moderator
99 Views

Hi MohnishJain,

This thread will no longer be monitored since we have provided a solution. If you need any additional information from Intel, please submit a new question.

 

 

Regards,

Hairul


Reply