Getting Label (classification) from predicted output (out)

[Boino]

Hello,

I have been trying to figure out how to extract the "name" of the object detected on the picture. I have come across the following issue:

When I run the simple Convnets code using the VGG_16 model with an image of a horse of size (4650x3260), I get an array "out(1, 1000)". I assumed that each index of the array represents an object (e.g. 85 = French Bulldog), and that each value of the array represents the probability of the image showing the class assigned to that index (e.g. 0.143~ 14.3% probability of Image containing a French Bulldog).

However, when I run the code and find out the highest value of the array "out", I get and index/id of 349. When I transform that index/id into synset, I get 'n02415577'. This all looked good until I checked the ImageNet classes here and found out that: a) Index/ID 349 does not correspond to synset n02415577, and b) neither the ID nor the synset are assigned to "horse" on ImageNet.

The code I have used to find out the Index ID is:

import numpy max_index = numpy.argmax(out)

and to find the synset:

from convnetskeras.imagenet_tool import id_to_synset wnid = id_to_synset(max_index)

Does anyone know why does it work fine for the given examples but performs so poorly with other images? Am I perhaps interpreting the prediction "out" completely wrong? Or does the image size play a significant role that could be causing problems?

Thanks

[sjwetzel]

Interpreting the images can be easily done with

https://gist.github.com/yrevar/942d3a0ac09ec9e5eb3a

you can also compare the results via

(out>0.03)[0].nonzero()

[loliwinston]

@Boino , did you find out the problem to this?