ann_vecs is dict of attribute vectors. The key is used to index into the patch_id_to_ann_id dict to get the ann id from the MS COCO annotation dataset (the original MS COCO dataset).
img_attrs = attr_data['ann_vecs'][idx]
ann_id = attr_data['patch_id_to_ann_id'][idx]
split is a dict which tells us whether a the image for the attribute vector is in the train split or val split.
Given the ann_id above, we now have to search for the same ann_id (given by id in the annotation dataset):
anns = [a for a in annotations['annotations'] if a['id'] == ann_id]
# anns should have only one entry since id is unique
ann = anns[0]
# Get the image name
img_id = ann['image_id']
To map the attributes in ann_vecs to the actual attributes, we need to sort the attributes list in attributes.
In the attributes dataset, ann_attrs is a dict where each key is the id (ann_id) of the annotation/object from the annotation dataset. The value for the ann_id will be a dictionary which has the following fields:
attrs_vectorwhich is the list of attribute values from the ELA algorithm.splitwhich tells us which split does the image belongs to.
The attributes list will be saved in the sorted order, sorted on id. This is the same order in which the attrs_vector is ordered.
Thus to get the attribute vector for any annotation would be:
ann = annotations['annotation'][idx] # idx is a valid index in the annotations dataset
attr = attributes['ann_attrs'][ann['id']]
attr_vector = attr['attrs_vector']
split = attr['split'] # this is partially redundant since our idx is obtained from either the train or val split