MOT20

MOT20 is a benchmark dataset for single-camera multi-object tracking (MOT) and pedestrian detection in very crowded real-world scenes. This Hugging Face repository provides MOT20 in the original MOTChallenge-style structure for research, benchmarking, training, and evaluation of multi-object tracking systems.

MOT20 was introduced to stress-test MOT methods in high-density pedestrian scenes, including crowded squares, indoor train stations, stadium exits, and pedestrian streets. Compared with earlier MOTChallenge releases, MOT20 contains fewer video sequences but substantially denser annotations.

This repository is a convenience mirror/repackaging of MOT20. Please cite the original MOT20 paper and follow the original MOTChallenge dataset terms and conditions.

Dataset Details

The official MOT20 benchmark contains 8 challenging video sequences: 4 training sequences and 4 test sequences. Tracking and evaluation are performed in image coordinates, and all sequences are annotated using the MOTChallenge protocol.

Supported Tasks

This dataset can be used for:

• Multiple Object Tracking (MOT)
• Multi-pedestrian tracking
• Tracking-by-detection research
• Pedestrian detection in crowded scenes
• Occlusion-heavy tracking research
• Re-identification-assisted tracking
• MOT benchmark conversion and evaluation pipelines

Dataset Structure

A typical MOT20 repository follows the official MOTChallenge folder layout:

MOT20/
├── train/
│   ├── MOT20-01/
│   ├── MOT20-02/
│   ├── MOT20-03/
│   └── MOT20-05/
├── test/
│   ├── MOT20-04/
│   ├── MOT20-06/
│   ├── MOT20-07/
│   └── MOT20-08/
└── README.md

Each sequence directory typically contains:

MOT20-XX/
├── img1/          # Video frames as image files
├── det/           # Public detections
│   └── det.txt
├── gt/            # Ground-truth annotations; training split only
│   └── gt.txt
└── seqinfo.ini    # Sequence metadata

Unlike MOT17, MOT20 sequences are not duplicated across DPM/FRCNN/SDP detector-specific folders. Each sequence is provided as a single MOTChallenge sequence folder, typically with public detections in det/det.txt.

Splits

Training Sequences

The MOT20 training split contains 4 video sequences:

Sequence	FPS	Resolution	Length	Tracks	Boxes	Density	Description
MOT20-01	25	1920x1080	429 frames	90	26,647	62.1	Crowded indoor train station
MOT20-02	25	1920x1080	2,782 frames	296	202,215	72.7	Crowded indoor train station
MOT20-03	25	1173x880	2,405 frames	735	356,728	148.3	People leaving a stadium entrance at night, elevated viewpoint
MOT20-05	25	1654x1080	3,315 frames	1,211	751,330	226.6	Crowded square at night

Training split totals:

• Frames: 8,931
• Duration: 357 seconds
• Tracks: 2,332
• Boxes: 1,336,920
• Average density: 149.7 boxes/frame

Test Sequences

The MOT20 test split contains 4 video sequences:

Sequence	FPS	Resolution	Length	Tracks	Boxes	Density	Description
MOT20-04	25	1545x1080	2,080 frames	728	371,525	178.6	Crowded square at night
MOT20-06	25	1920x734	1,008 frames	368	207,543	205.9	Pedestrian street scene
MOT20-07	25	1920x1080	585 frames	126	41,096	70.2	Crowded indoor train station
MOT20-08	25	1920x734	806 frames	279	145,301	180.3	Pedestrian street scene

Test split totals:

• Frames: 4,479
• Duration: 178 seconds
• Tracks: 1,501
• Boxes: 765,465
• Average density: 170.9 boxes/frame

Test-set ground truth is not included in the public release. Official evaluation should be done through the MOTChallenge platform.

Annotation Format

MOT20 uses the standard MOTChallenge comma-separated text format.

Ground Truth Format

Training annotations are stored in:

gt/gt.txt

Each row generally follows:

<frame>, <id>, <bb_left>, <bb_top>, <bb_width>, <bb_height>, <conf>, <class>, <visibility>

Field descriptions:

Field	Description
frame	Frame index, starting from 1
id	Object identity ID
bb_left	Left coordinate of bounding box
bb_top	Top coordinate of bounding box
bb_width	Bounding-box width
bb_height	Bounding-box height
conf	Confidence flag for ground truth
class	Object class label
visibility	Visibility ratio / visibility flag

Detection Format

Public detections are stored in:

det/det.txt

Each row generally follows:

<frame>, <id>, <bb_left>, <bb_top>, <bb_width>, <bb_height>, <conf>, <x>, <y>, <z>

For detection files, id is commonly set to -1, and the final world-coordinate fields may be unused depending on the detector or sequence.

Usage

Download from Hugging Face

from huggingface_hub import snapshot_download

repo_dir = snapshot_download(
    repo_id="YOUR_USERNAME_OR_ORG/MOT20",
    repo_type="dataset",
)

print(repo_dir)

Replace YOUR_USERNAME_OR_ORG/MOT20 with the actual Hugging Face dataset repository ID.

Example: Read MOTChallenge Annotations with Python

from pathlib import Path
import pandas as pd

seq_dir = Path("MOT20/train/MOT20-01")
gt_path = seq_dir / "gt" / "gt.txt"

gt = pd.read_csv(
    gt_path,
    header=None,
    names=[
        "frame",
        "id",
        "bb_left",
        "bb_top",
        "bb_width",
        "bb_height",
        "conf",
        "class",
        "visibility",
    ],
)

print(gt.head())

Example: Iterate Over Frames

from pathlib import Path

img_dir = Path("MOT20/train/MOT20-01/img1")
frames = sorted(img_dir.glob("*.jpg"))

print(f"Number of frames: {len(frames)}")
print(frames[:5])

Example: Read Public Detections

from pathlib import Path
import pandas as pd

seq_dir = Path("MOT20/train/MOT20-01")
det_path = seq_dir / "det" / "det.txt"

det = pd.read_csv(
    det_path,
    header=None,
    names=[
        "frame",
        "id",
        "bb_left",
        "bb_top",
        "bb_width",
        "bb_height",
        "conf",
        "x",
        "y",
        "z",
    ],
)

print(det.head())

Evaluation

For official MOTChallenge-style evaluation, use the MOTChallenge evaluation protocol and compatible tools such as TrackEval.

Typical predicted tracking result format:

<frame>, <id>, <bb_left>, <bb_top>, <bb_width>, <bb_height>, <conf>, <x>, <y>, <z>

Common MOT metrics include:

• MOTA
• MOTP
• IDF1
• HOTA
• FP / FN / ID switches
• Mostly Tracked / Mostly Lost trajectories

Official benchmark submissions should be made through the MOTChallenge platform, not through this Hugging Face repository.

Intended Use

This dataset is intended for:

• Academic research in multi-object tracking
• Benchmarking MOT algorithms in very crowded scenes
• Studying occlusion, dense pedestrian motion, and detector behavior under crowding
• Developing tracking-by-detection pipelines
• Training and validating pedestrian tracking systems

Limitations and Responsible Use

MOT20 contains real-world pedestrian scenes from public or semi-public environments. Users should consider privacy, surveillance, and fairness implications when training or deploying models using this dataset.

Known limitations include:

• The dataset focuses on dense pedestrian scenes and is not representative of all tracking scenarios.
• Test-set ground truth is not included in the public dataset release.
• High crowd density and occlusion make identity preservation especially difficult.
• The dataset is not representative of all countries, camera types, lighting conditions, or pedestrian demographics.
• Models trained on MOT20 should be evaluated carefully before deployment in real-world surveillance or safety-critical systems.

Licensing and Redistribution

The license metadata for this Hugging Face dataset card is set to:

license: cc-by-nc-sa-3.0

The MOTChallenge datasets are published under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License. Please refer to the original MOTChallenge website and competition pages for the authoritative dataset access terms, redistribution rules, and citation requirements. Do not assume this mirror grants rights beyond those provided by the original dataset owners.

Citation

Please cite the MOT20 benchmark paper if you use this dataset:

@article{dendorfer2020mot20,
  title={MOT20: A benchmark for multi object tracking in crowded scenes},
  author={Dendorfer, Patrick and Rezatofighi, Hamid and Milan, Anton and Shi, Javen and Cremers, Daniel and Reid, Ian and Roth, Stefan and Schindler, Konrad and Leal-Taixe, Laura},
  journal={arXiv preprint arXiv:2003.09003},
  year={2020}
}

You may also cite the general MOTChallenge benchmark paper when appropriate:

@article{dendorfer2020motchallenge,
  title={MOTChallenge: A Benchmark for Single-Camera Multiple Target Tracking},
  author={Dendorfer, Patrick and Osep, Aljosa and Milan, Anton and Schindler, Konrad and Cremers, Daniel and Reid, Ian and Roth, Stefan and Leal-Taixe, Laura},
  journal={International Journal of Computer Vision},
  year={2020},
  doi={10.1007/s11263-020-01393-0}
}

References

• MOTChallenge MOT20: https://motchallenge.net/data/MOT20/
• MOT20 paper: https://arxiv.org/abs/2003.09003
• MOTChallenge website: https://motchallenge.net/
• TrackEval evaluation toolkit: https://github.com/JonathonLuiten/TrackEval