The direct download link for the data can be found here https://www.cs.cornell.edu/projects/field-guide/static/data/data.zip.
The data directory should be extracted in the home directory of the project.

So if `path/to/Field-Guide-ZSL` is the project directory data directory should be `path/to/Field-Guide-ZSL/data`.

The data directory is organized in the following format:

data/
    CUB/
        CUB_r101.pkl
        CUB_parent.npy
        CUB_nns.npy
    SUN/
       ...
    AWA2/
       ...    

`{Dataset}_r101.pkl` contains information semantic and image information for the dataset. The `.pkl` file contains the information provided by the datasets from [1]. 

`{Dataset}_parent.npy` contains the taxonomy information about the dataset. Classes with same parents are considered siblings in our framework.

`{Dataset}_nns.npy` contains the user annotated nearest base class information for the novel classes.

[1] Zero-Shot Learning - A Comprehensive Evaluation of the Good, the Bad and the Ugly, Xian et. al., CVPR 2017


-----------------------------------------------------------------------------------------

Description of '{Dataset}_r101.pkl'
This is a pickle file containing following keys:
features : 2-d numpy array (size: Nx2048) containing resnet features for all N images (train/test/val).
labels: 1-d numpy array (size: N) containing class labels for all N images (train/test/val).
class_names: 1-d numpy array (size: C) containing class names corresponding to index for all C classes (seen/unseen).
image_files: 1-d numpy array (size: N) containing file name for all N images (train/test/val). We do not use this in our experiements but this could be useful for future analysis.
attributes: 2-d numpy array (size: CxA) containing normalized attribute description of length A for C classes (seen/unseen).
attribute_names: python list (length: A) containing name of the A attributes.
original_attributes: 2-d numpy array (size: CxA) containing original attribute description of length A for C classes (seen/unseen).
trainval_loc: 1-d numpy array containing indices of image-labels used for training and validation (training for deployment).
train_loc: 1-d numpy array containing indices of image-labels used for training.
val_loc: 1-d numpy array containing indices of image-labels used for validation.
test_seen_loc: 1-d numpy array containing indices of image-labels used for testing containing classes seen during training.
test_unseen_loc: 1-d numpy array containing indices of image-labels used for testing containing classes not seen during training.
perimageatrribute: Dictionary (size: N) containing image-level attributes for each image annotated by humans. These attributes are aggregated to get class level attributes. Values that are not use NaN instead.