Hi, I'm hyunwoo lee
I'm using the movidius and raspberry pi to recognize classify images.
But after I convert caffe model to movidius graph , I got the different result.
So I checked same image for two models ( one is caffe model , the other is movidius model)
I used Resnet-18 model and my target class is 5.
As you can see , result is so different .
CAFFE / resnet18_deploy / resnet18_custom_transfer_iter_5000
[0.15182055201006436, 0.15381028087897108, 0.3939189286761392, 0.15019196295911535, 0.15025827547570988]
CAFFE / resnet18_deploy / resnet18_custom_transfer_v2_iter_4000
[0.2008975630376957, 0.1603913024819096, 0.19200574522089883, 0.16627960206785145, 0.2804257871916444]
MOVIDIUS / resnet18_custom_transfer_iter_5000.graph
[ 4.99725342e-03 2.81494141e-01 7.12890625e-01 0.00000000e+00 5.64575195e-04]
MOVIDIUS / resnet18_custom_transfer_v2_iter_4000.graph
[ 9.86328125e-01 8.58306885e-04 5.98144531e-03 5.03921509e-03 1.46579742e-03]
@hyunwoolee Hi, how did you check your ResNet model on the NCS? Can you provide your NCS code and ResNet 18 model for testing? With the NCS, you may have to perform some manual image preprocessing (mean subtraction and scaling) before inference. WIth Caffe, the preprocessing is done automatically.