Awesome

Youtube-Infant-Body-Parsing Video Dataset

The infant video dataset collected from Youtube with body parsing annotations, which is used in our Medical Image Analysis 2023 paper Semi-supervised body parsing and pose estimation for enhancing infant general movement assessment. The data and annotations are released for research purposes only.

Dataset Description

This dataset includes 90 infant movement videos collected from Youtube by Chambers et al. (2020). URLs of videos are provided in their github. We downloaded videos and downsampled them every 4 frames due to the long length of original videos. Sampled frames are available on this google drive. The format of each frame name is {7_digits_of_video_index}{6_digits_of_frame_index}. For infant body parsing, we collect annotations for five classes: background, head, arm, torso and leg. We randomly split all videos into 68 training videos and 22 testing videos, resulting in 2,149 labeled and 4,690 unlabeled training frames, and 1,256 labeled and 2,737 unlabeled testing frames. They corresponds to train_label.json, train_unlabel.json, test_label.json, and test_unlabel.json, respectively. The following codes show how to use these json files. Note that the data and annotations are released for research purposes only.

import json
import base64
import numpy as np
import cv2

def base64tonpy(file):
    base64_decoded = base64.b64decode(file)
    im_arr = np.frombuffer(base64_decoded, dtype=np.uint8)
    img = cv2.imdecode(im_arr, flags=cv2.IMREAD_COLOR)
    return img[:, :, ::-1]
    
def im2vl(img):
    img_tmp = np.zeros(img.shape[:2], dtype=np.uint8)
    head_msk = np.all(img == [44, 160, 44], axis=2)
    hand_msk = np.all(img == [31, 119, 180], axis=2)
    body_msk = np.all(img == [255, 127, 14], axis=2)
    foot_msk = np.all(img == [214, 39, 40], axis=2)
    img_tmp[head_msk] = 1
    img_tmp[hand_msk] = 2
    img_tmp[body_msk] = 3
    img_tmp[foot_msk] = 4
    return img_tmp

train_label_path = "train_label.json"
train_unlabel_path = "train_unlabel.json"
test_label_path = "test_label.json"
test_unlabel_path = "test_unlabel.json"

# here we use train_label and train_unlabel as examples
with open(train_label_path, "r") as f:
    train_label_dict = json.load(f)
with open(train_unlabel_path, "r") as f:
    train_unlabel_dict = json.load(f)

# each key of dict is the video name
print(list(train_label_dict.keys())[0:3])
# ['1000000', '1000002', '1000004']

# each item of dict contains a frame dict list
# each frame dict contains "image" (and "label")
print(train_label_dict["1000002"][0:3])
# [{'image': '1000002000000.jpg', 'label': ...}, {'image': '1000002000012.jpg', 'label': ...}]

# change label from base64 to numpy format
example_label = train_label_dict["1000002"][0]['label']
example_label = base64tonpy(example_label)
# example label is RGB label image
# background [0, 0, 0]
# head [44, 160, 44]
# hand [31, 119, 180]
# body [255, 127, 14]
# foot [214, 39, 40]
print(example_label.shape)
# (360, 640, 3)
# change to normal label image
example_label = im2vl(example_label)
print(example_label.shape)
# (360, 640)

Citing Youtube-Infant-Body-Parsing

If you find this dataset useful in your research, please consider citing our paper (collecting anatotations) and Chambers' paper (collecting videos):

@article{ni2023semi,
  title={Semi-supervised body parsing and pose estimation for enhancing infant general movement assessment},
  author={Ni, Haomiao and Xue, Yuan and Ma, Liya and Zhang, Qian and Li, Xiaoye and Huang, Sharon X},
  journal={Medical Image Analysis},
  volume={83},
  pages={102654},
  year={2023},
  publisher={Elsevier}
}

@article{chambers2020computer,
  title={Computer vision to automatically assess infant neuromotor risk},
  author={Chambers, Claire and Seethapathi, Nidhi and Saluja, Rachit and Loeb, Helen and Pierce, Samuel R and Bogen, Daniel K and Prosser, Laura and Johnson, Michelle J and Kording, Konrad P},
  journal={IEEE Transactions on Neural Systems and Rehabilitation Engineering},
  volume={28},
  number={11},
  pages={2431--2442},
  year={2020},
  publisher={IEEE}
}

For any problems with the data and annotations, please feel free to open an issue or contact me: homerhm.ni@gmail.com

Acknowledgement

The original Youtube videos are collected by Chambers et al. (2020) and the labels are annotated through the Amazon Mechanical Turk crowd-sourcing platform. We also specially thank Yifei Xiang from the EE Department of Penn State University for his significant role in data pre-processing and annotation.