JTA (Joint Track Auto) is a huge dataset for pedestrian pose estimation and tracking in urban scenarios created by exploiting the highly photorealistic video game Grand Theft Auto V. We collected a set of 512 full-HD videos (256 for training and 256 for testing), 30 seconds long, recorded at 30 fps. The dataset was created with this tool.

You can download JTA here. By downloading the dataset you agree on the following statement: "I declare that I will use the JTA Dataset for research and educational purposes only, since I am aware that commercial use is prohibited. I also undertake to purchase a copy of Grand Theft Auto V."

After the data download, your JTA-Dataset directory will contain the following files:

• annotations: directory with dataset annotations

  • annotations/train: directory with 256 JSON files, one for each training sequence
  • annotations/test: directory with 128 JSON files, one for each testing sequence
  • annotations/val: directory with 128 JSON files, one for each validation sequence

• videos: directory with dataset videos

  • videos/train: directory with 256 videos (.MP4), one for each training sequence
  • videos/test: directory with 128 videos (.MP4), one for each testing sequence
  • videos/val: directory with 128 videos (.MP4), one for each validation sequence

• to_imgs.py: Python script that splits the videos into its frames and saves them in a specified directory with the desired format (default = JPG)

  • requires Python >= 3.6 (see requirements.txt for more details)
  • use python to_imgs.py --help to read the help message
  • use option --img_format='png' for better quality
  • each frame has a size of 1920×1080 px
  • usage example:

    python to_imgs.py --out_dir_path='frames' --img_format='jpg'

• to_poses.py: Python script that splits the per sequence annotations into per frame annotations and saves them in a specified directory with the desired format (default = numpy)

  • requires Python >= 3.6 (see requirements.txt for more details)
  • use python to_poses.py --help to read the help message
  • use option --format=torch to save the annotations using torch
  • usage example:

    python to_poses.py --out_dir_path='poses' --format='numpy'

  WARNING: before loading those annotations you have to import Joint from joint.py and Pose from pose.py.

• visualize.py: Python script that provides a visual representation of the annotations.

  • requires Python >= 3.6 (see requirements.txt for more details)
  • use python visualize.py --help to read the help message
  • usage example:

    python visualize.py --in_mp4_file_path='videos/train/seq_42.mp4' --json_file_path='annotations/train/seq_42.json' --out_mp4_file_path='vis_ann/seq_42.mp4'

• coco_style_convert.py: Python script for annotation conversion (from JTA format to COCO format).

  • requires Python >= 3.6 (see requirements.txt for more details)
  • use python coco_style_convert.py --help to read the help message
  • usage example:

    python coco_style_convert.py --out_dir_path='coco_annotations'

• posetrack_style_convert.py: Python script for annotation conversion (from JTA format to PoseTrack18 format).

  • requires Python >= 3.6 (see requirements.txt for more details)
  • use python posetrack_style_convert.py --help to read the help message
  • usage example:

    python posetrack_style_convert.py --out_dir_path='posetrack_annotations --max_frame=200'

• joint.py and pose.py: support classes for the scripts.

Each annotation file refers to a specific sequence (e.g. seq_42.json is the annotation file of seq_42.mp4). An annotation file consists of a list of lists, that is a matrix having N rows and 10 columns. Each row of the matrix contains the data of a joint; these data are organized as follows:

• Note #1: 2D coordinates are relative to the top left corner of the frame, while 3D coordinates are in the standard camera coordinate system.
• Note #2: frames are counted starting from 1.

Each sequence has been recorded with the same camera with the followng intrinsic matrix:

The associations between numerical identifier and type of joint are the following:

 0: head_top
 1: head_center
 2: neck
 3: right_clavicle
 4: right_shoulder
 5: right_elbow
 6: right_wrist
 7: left_clavicle
 8: left_shoulder
 9: left_elbow
10: left_wrist
11: spine0
12: spine1
13: spine2
14: spine3
15: spine4
16: right_hip
17: right_knee
18: right_ankle
19: left_hip
20: left_knee
21: left_ankle

• Read data of an annotation file and convert it to numpy array:

  import json
  import numpy as np
  
  json_file_path = 'annotations/training/seq_42.json'
  
  with open(json_file_path, 'r') as json_file:
      matrix = json.load(json_file)
      matrix = np.array(matrix)

• Get data of frame #118:

  frame_data = matrix[matrix[:, 0] == 118]

• Get data of all the joints of person with ID=3 in frame #118:

  person_data = frame_data[frame_data[:, 1] == 3]

• Note: we have chosen to save our annotations in JSON format for portability reasons. However, if you want to speed up data loading in Python, we suggest converting the annotations to numpy format:

  np.save(json_file_path.replace('json', 'npy'), matrix)
  

If you want, you can convert our annotations to COCO format using the coco_style_convert.py script, but note that occlusion, tracking and 3D informations are not available in that format.

This dataset was introduced in the paper "Learning to Detect and Track Visible and Occluded Body Joints in a Virtual World" (ECCV 2018). In the experimental section of the paper, when referring to the "test set", we mean the set consisting of the test and val directories of the JTA Dataset.

We believe in open research and we are happy if you find this data useful.
If you use it, please cite our work.

@inproceedings{fabbri2018learning,
   title     = {Learning to Detect and Track Visible and Occluded Body Joints in a Virtual World},
   author    = {Fabbri, Matteo and Lanzi, Fabio and Calderara, Simone and Palazzi, Andrea and Vezzani, Roberto and Cucchiara, Rita},
   booktitle = {European Conference on Computer Vision (ECCV)},
   year      = {2018}
 }

JTA-Dataset is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.