The convolution operation is different in tensorflow and caffe, especially the padding iterm. Thus, using directly converted model will cause problems. Trick: using a pad = 2 convolution operation for CONV1 Layer to get a larger feature map and than crop into the right size.
But for kernel size 3 x 3 or 1 x 1 and stride step 1 with padding SAME in tensorflow, we directly choose convolution param with pad = 1, kernel size = 3 and stride = 1
cd caffe-squeezeDet
python ./src/demo.py