MTA

Dataset Overview

Basic Information

• Name: MTA (Multi Camera Track Auto)
• Type: Large multi‑target multi‑camera tracking dataset
• Scale: Contains over 2,800 person identities, captured by 6 cameras, each video longer than 100 minutes
• Temporal Coverage: Includes both daytime and nighttime periods

Dataset Content

Multi‑Person Multi‑Camera Tracking

• Video files:

  • MTA_videos.zip: Contains 12 videos (training and testing videos for 6 cameras), total length 102 minutes, 41 GB compressed, 42 GB uncompressed
  • MTA_videos_coords.zip: Contains full annotations, 28.6 GB compressed, 235 GB uncompressed
  • MTA_ext_short.zip: Contains extracted short video segments and annotations, 1.7 GB compressed, 1.8 GB uncompressed, total length 4 minutes
  • MTA_ext_short_coords.zip: Contains full annotations for the extracted short segments, 1.1 GB compressed, 8.9 GB uncompressed

• Annotation files:

  • coords_fib_cam_{0-5}.csv: Frame number, person ID, bounding‑box coordinates
  • coords_cam_{0-5}.csv: Frame number, person ID, appearance ID, joint type, 2D/3D joint positions, joint occlusion status, and other detailed information

Person Re‑Identification

• Dataset: MTA_reid.zip, 0.8 GB compressed, 1.2 GB uncompressed

  • Training set: 72,301 images
  • Query set: 15,165 images
  • Test set: 60,448 images
  • Distractor images: 36,100 images, lacking sufficient person identification information

• Annotations: Embedded in image filenames, e.g., framegta_{int}_camid_{0-5}_pid_{int}.png, providing in‑game frame number, camera ID, and person ID

Dataset Access

• Must contact the dataset providers via email to obtain download links, providing personal information and the intended purpose of data usage.

Citation Information

• When citing this dataset, refer to the related paper presented at the CVPR 2020 VUHCS Workshop.

The above information provides a concise description of the MTA dataset, covering basic information, content composition, acquisition method, and citation requirements.