ILSVRC2012 validation ground_truth labels index not matching pretrained networks index

Question

I'm currently using VGG-S pretrained convolutional neural network provided by Lasagne library, from the following link.

I downloaded the validation dataset of ILSVRC2012 in order to do a classification task, but I saw that the ground truth values found in the file ILSVRC2012_devkit_t12_v2/data/ILSVRC2012_validation_ground_truth.txt didn't match at all the output values of the VGG-S network (softmax values).

For example: the last image ILSVRC2012_val_00050000.JPEG which is clearly a Llama, vgg-s outputs label n°355 (llama), while the ground truth label found in the file is n° 186 (Norwich terrier).

The index-to-label is the same as this one.

Can anyone please tell me where I'm wrong ? Thanks for the help.

Papert

Answer 1

The ground truth system you are using is developed by caffe group, where is different from Imagenet standard groundtruth indexing system.

https://github.com/BVLC/caffe/tree/master/data/ilsvrc12

Replace your groungtruth with val.txt.

Peshawar

Answer 2

Never mind, I found out that VGG convolutional networks used a different class order according to this note

The validation ground truth labels that were used can be downloaded from here

Papert