KRadim/AwA-Pose-Lite

Dataset Summary

AwA Pose, a quadrupedal keypoint detection dataset that offers richer annotations and greater species diversity than existing datasets. The data supports efficiency, advances in research on generalized keypoint detection in animals.

The dataset was explored, filtered, and stored in a pyarrow type. See : AwA2_dataset_analysis

Original Kaggle dataset: AwA2_dataset

The original data comes from two sources and is described in:

• A Novel Dataset for Keypoint Detection of Quadruped Animals from Images
• GitHub: prinik/AwA-Pose
• Animals with Attributes 2

The dataset contains:

• Lite version of keypoints dataset
• Full version of keypoints dataset

All version contains:

• train dataset: 90%
• validation dataset: 5%
• test dataset: 5%

Description of data in the dataset

Class names:

• antelope
• bobcat
• buffalo
• chihuahua
• collie
• cow
• dalmatian
• deer
• elephant
• fox
• german+shepherd
• giant+panda
• giraffe
• grizzly+bear
• hippopotamus
• horse
• leopard
• lion
• moose
• otter
• ox
• persian+cat
• pig
• polar+bear
• rabbit
• raccoon
• rhinoceros
• sheep
• siamese+cat
• squirrel
• tiger
• weasel
• wolf
• zebra

Columns description

name column	description of column
id	id number of records
right_eye	keypoint values [x,y]
right_earbase	keypoint values [x,y]
right_earend	keypoint values [x,y]
right_antler_base	keypoint values [x,y]
right_antler_end	keypoint values [x,y]
left_antler_base	keypoint values [x,y]
left_antler_end	keypoint values [x,y]
left_earbase	keypoint values [x,y]
left_earend	keypoint values [x,y]
left_eye	keypoint values [x,y]
nose	keypoint values [x,y]
upper_jaw	keypoint values [x,y]
lower_jaw	keypoint values [x,y]
mouth_end_right	keypoint values [x,y]
throat_base	keypoint values [x,y]
neck_base	keypoint values [x,y]
neck_end	keypoint values [x,y]
back_base	keypoint values [x,y]
back_middle	keypoint values [x,y]
back_end	keypoint values [x,y]
tail_base	keypoint values [x,y]
body_middle_right	keypoint values [x,y]
bbox	bounding box dimension [x1, y1, x2, y2]
mouth_end_left	keypoint values [x,y]
throat_end	keypoint values [x,y]
tail_end	keypoint values [x,y]
front_left_thai	keypoint values [x,y]
front_left_knee	keypoint values [x,y]
front_left_paw	keypoint values [x,y]
front_right_thai	keypoint values [x,y]
front_right_paw	keypoint values [x,y]
front_right_knee	keypoint values [x,y]
back_left_knee	keypoint values [x,y]
back_left_paw	keypoint values [x,y]
back_left_thai	keypoint values [x,y]
back_right_thai	keypoint values [x,y]
back_right_paw	keypoint values [x,y]
back_right_knee	keypoint values [x,y]
belly_bottom	keypoint values [x,y]
body_middle_left	keypoint values [x,y]
name_file	name of file as string
name_class	name of class as string
image_base64s	image as string Base64
image_width	value of with of image
image_height	value of height of image
image_license	text of licence of image

Notes on data

If a keypoints contains [-1.0, -1.0], it means that the point is not visible in the image. These points must be masked when training the model.

Images are stored as a Base64 string. They can be transformed using the function:

import base64
import io
from PIL import Image

def base64_to_img(base64_str):
    img_bytes = base64.b64decode(base64_str)
    img_buffer = io.BytesIO(img_bytes)
    img = Image.open(img_buffer)
    
    return img

Find out more in AwA2_dataset_analysis

Licensing Information

Data for keypoints is licensed according to GitHub: prinik/AwA-Pose, this license is MIT.

The license for the images is according to Animals with Attributes 2, see data column image_license .