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.
5922 Discussions

Compiling MobileNetSSD produce wrong results

idata
Community Manager
366 Views

I followed the instructions in this post to compile a MobileNetSSD fine tuned on my own dataset using NCSDK1. I can compile the model generated by merge_bn.py. However, the results are not good. This is what I tried:

 

     

  • original merge_bn script, original deploy.prototxt: good results with cpu but mvNCCompile gives error (as expected)
  •  

  • original merge_bn script, modified deploy.prototxt (without "_new" parts): bad results both with cpu and movidius
  •  

  • modified merge_bn script (according to the mentioned post), original prototxt: error in merge_bn (as expected)
  •  

  • modified merge_bn script, modified deploy_prototxt (as suggested by the post): bad results both with cpu and movidius.
  •  

 

Does anyone have some suggestion about how to fix this problem?

0 Kudos
1 Reply
idata
Community Manager
136 Views

Update: I think I fixed the problem not by removing the "new" part from the layer name, but adding "new" to the "top" of the layer and to every reference to the layer.

Reply