Attribute Description Dataset of Animal(ADDA)

Animal Feature Relation Data

The animal list of this data is based on Animals with Attributes 2. We collected animal feature explanations by googling.

We call this data as "animal relation data" because we are considering the relations of words in the sentence that explains animal.

animal relation data distribute of the data

• Animal Number : 50 (which is the same as number of animals in AWA data)
• Relation Number : 48
• Category Number : 4

What is Category?

We split explanations of animals into 4 categories(appearance, habitat, diet, characteristic). These categories are a large classification of the feature found in animal images. additionally, category characteristic is for the features that cannot be included in any other categories

Overview

Detailed Explanation of Data

subj and obj types

subj count: 39 obj count: 37

You can check the detailed list of subj and obj type in subj_obj_type.json There is special subj type named KEY_SUBJ, which means the subj is not in the sentence, but you can find it in key value of the sentence. for example,

	"otter":[
        {
            "sentence": [
                "small",
                "mammals",
                "may",
                "also",
                "be",
                "eaten",
                "."
            ],
            "relation": [
                {
                    "category": "diet",
                    "subj_st": -1,
                    "subj_end": -1,
                    "subj_type": "KEY_SUBJ",
                    "obj_st": 0,
                    "obj_end": 1,
                    "obj_type": "food",
                    "rel_type": "animal_feed"
                }
            ]
        }
    ]

in this sentence, there is no word that indicate animal otter. So we chose subj_type as KEY_SUBJ and express word idx(subj_st, subj_end) as -1.

Relations which include less than 7 animals.

"stride_is": 1, "animal_love": 1, "wing_is": 1, "bigger_than": 1, "larger_than": 1, "smaller_than": 1, "chest_is": 1, "shoulder_is": 1, "baleen_is": 2, "group_is": 2, "longer_than": 2, "finger_is": 2, "flipper_is": 4, "color_is": 5, "hump_is": 6, "claw_is": 6, "hand_is": 7

To consider all relations in training, I think it would be better to include the following animals in the study.

22 animals ['antelope', 'hippopotamus', 'bat', 'rat', 'buffalo', 'mouse', 'collie', 'blue+whale', 'chimpanzee', 'giraffe', 'giant+panda', 'raccoon', 'walrus', 'dolphin', 'fox', 'bobcat', 'grizzly+bear', 'moose', 'skunk', 'otter', 'mole', 'seal']

But you can change anyway.