segmentation_sample get different output with same image

[junyi]

when I was experiencing with segmentation_sample, the result is abnormal when the batch size is greater than one

inside /tmp/test folder got two same image (18_38_55.jpg)

run the segmentation_sample with following command

~/intel/computer_vision_sdk/inference_engine/samples/intel64/Release/segmentation_sample -m ~/model.xml -i /tmp/test

the expect result is out_0.bmp should equal to out_1.bmp, instead of two different image

[Monique_J_Intel]

Hi Junyi,

Can you give me more information on the model that you are using so that I can reproduce your issue? Also, Is the input image that you have attached purposefully all black?

Kind Regards,

Monique Jones

[junyi]

Hi Monique,

Due to confidentiality reason, i am not allowed to share information/model i am working on.  However, like to know if you or the community has experienced problem with batch size > 1 on inference on CPU for segmentation.  Or is there any tip you can give us for us to try?  One thing worth noting is that for the same model, we are able to inference with batch size > 1 on Movidius, but are only able to with only batch size = 1 with CPU(i5).  This leads us to suspect that the issue may have something to do with how hardware handle the batch.  Any tip is much appreciated.

Thanks,

[Monique_J_Intel]

Hi Junyi,

I understand your confidentiality reasoning, I will try to reproduce this behavior with a public FCN8 Segmentation model tomorrow and respond back to you hopefully with some insight on this issue.

Kind Regards,

Monique Jones

[Monique_J_Intel · Accepted solution]

Hi Junyi,

I tested your scenario with an FCN8 model and couldn't reproduce your issue. I got the expected behavior of getting two images that are fully purple. Also, during Model Optimizer stage I did set the -b flag to 2 and you may want to try this? Also are you using a FCN8 model?

Kind Regards,

Monique Jones

[junyi]

Hi Monique,

I appreciated your time and effort, but due to confidentiality reason, I'm not allowed to share anything future information about the model that I'm working on.
And I did set the batch size to 2 by using -b flag. Also, I'm really glad to know that you got the expected behavior by using FCN8 model, which gives me a way that our model might cause the problem

Thanks,