Caffe Classification C++ Example vs Digits Python inference.py

[Luiz Claudio Silva Magnago]

I'm having better results with Digits than the caffe classification c++ example. I've heard that the difference is in preprocess, but I don't know exactly what is that difference. Someone has solved this problem? Someone has a good C++ classication code with the same preprocess that Digits has?

Thanks!

[Steve Herrick]

I am having the same issue. I can get DIGITS to classify my test images correctly however when I download the data and run the classification C++ code I get the wrong answers. I tried the classify.py script and it now complains about the dimensions of my mean.binaryproto file. I used 256 x 256 images but the python script is saying it doesn't match the 224 x 224 input. I don't know where it is getting 224 x 224 from.

[johnny b]

I have exactly the same issue! From my point of view this is a big downside of DIGITS that it is kind of opaque what kind of preprocessing steps it is using.
 
Would it not be possible to output some kind of preprocessing configuration file making it possible to preprocess the images according to this configuration?

[Greg Heinrich]

Hello,

> this is a big downside of DIGITS that it is kind of opaque what kind of preprocessing steps it is using

This is open-source software. If you can read code, none of this should be opaque.

[johnny b]

Am Montag, 9. Januar 2017 15:14:26 UTC+1 schrieb Greg Heinrich:

This is open-source software. If you can read code, none of this should be opaque.

That is true and it is nice that it is open source. Just saying that it would be a nice feature. E.g. if you click on "download model" you get not only the Caffe files, you also get a file which tells you how you need to preprocess your images before passing them to the network.

Currently I get almost the same result in C++ compared to DIGITS "Classify One Image" but somehow different results using "use_archive.py" and I set "Squash" as the images resizing method and mean pixel substracting while creating the model.