SportsMOT

SportsMOT is a large-scale dataset for single-camera multi-object tracking (MOT) in sports videos. It focuses on tracking players in professional sports scenes, where targets exhibit fast and variable-speed motion, frequent occlusion, motion blur, camera motion, and similar team uniforms.

This Hugging Face repository provides SportsMOT in a MOTChallenge-style structure for research, benchmarking, training, and evaluation of multi-object tracking systems in sports analytics.

This repository is a convenience mirror/repackaging of SportsMOT. Please cite the original SportsMOT paper and follow the original dataset license and terms.

Dataset Details

SportsMOT was introduced for MOT in sports scenes, where the objective is to track players on the playground/court. Spectators, referees, coaches, balls, and other non-player objects are generally not the tracking targets.

The dataset contains 240 sports video clips from basketball, football/soccer, and volleyball. The official paper reports over 150K frames, over 1.6M bounding boxes, and 3,401 tracks. SportsMOT is designed to stress-test object association, especially under fast motion, camera movement, occlusion, and visually similar athletes.

Supported Tasks

This dataset can be used for:

• Multiple Object Tracking (MOT)
• Player tracking in sports videos
• Athlete detection
• Tracking-by-detection research
• Appearance-based and motion-based association research
• Re-identification-assisted tracking
• Sports video analytics
• MOT benchmark conversion and evaluation pipelines

Dataset Structure

A typical SportsMOT repository follows a MOTChallenge-style layout with additional split metadata and helper scripts:

SportsMOT/
├── splits_txt/
│   ├── basketball.txt
│   ├── volleyball.txt
│   ├── football.txt
│   ├── train.txt
│   ├── val.txt
│   └── test.txt
├── scripts/
│   ├── mot_to_coco.py
│   └── sportsmot_to_trackeval.py
├── dataset/
│   ├── train/
│   │   ├── VIDEO_NAME_1/
│   │   │   ├── gt/
│   │   │   │   └── gt.txt
│   │   │   ├── img1/
│   │   │   │   ├── 000001.jpg
│   │   │   │   ├── 000002.jpg
│   │   │   │   └── ...
│   │   │   └── seqinfo.ini
│   │   └── ...
│   ├── val/
│   │   └── ...
│   └── test/
│       ├── VIDEO_NAME_1/
│       │   ├── img1/
│       │   │   ├── 000001.jpg
│       │   │   ├── 000002.jpg
│       │   │   └── ...
│       │   └── seqinfo.ini
│       └── ...
└── README.md

Training and validation sequences contain ground-truth annotations. Test sequences are intended for benchmark evaluation and may not include public ground-truth annotations.

Each sequence directory typically contains:

VIDEO_NAME/
├── img1/          # Video frames as image files
├── gt/            # Ground-truth annotations; train/val splits only
│   └── gt.txt
└── seqinfo.ini    # Sequence metadata

A typical seqinfo.ini file contains fields such as:

[Sequence]
name=v_-6Os86HzwCs_c001
imDir=img1
frameRate=25
seqLength=825
imWidth=1280
imHeight=720
imExt=.jpg

Splits

SportsMOT is split into training, validation, and test subsets:

Split	Number of videos	Ground truth
train	45	Included
val	45	Included
test	150	Usually withheld for benchmark evaluation
Total	240	—

The official split files are provided under splits_txt/:

splits_txt/train.txt
splits_txt/val.txt
splits_txt/test.txt
splits_txt/basketball.txt
splits_txt/volleyball.txt
splits_txt/football.txt

These text files map video sequence names to either dataset splits or sports categories.

Dataset Statistics

The original SportsMOT paper reports the following overall dataset statistics:

Statistic	Value
Video sequences	240
Frames	150,379
Duration	6,015 seconds
Bounding boxes	1,629,490
Tracks	3,401
Resolution	720p
Frame rate	25 FPS

Category-level statistics reported by the paper are summarized below. These values are category averages reported by the authors.

Category	Avg. frames	Avg. tracks	Avg. track gap	Avg. track length	Avg. boxes / frame
Basketball	845.4	10.0	68.7	767.9	9.1
Volleyball	360.4	12.0	38.2	335.9	11.2
Football / Soccer	673.9	20.5	116.1	422.1	12.8
Total	626.6	14.2	96.6	479.1	10.8

Annotation Format

SportsMOT uses a MOTChallenge-style comma-separated annotation format.

Ground Truth Format

Training and validation annotations are stored in:

gt/gt.txt

Each row follows the standard MOT-style ground-truth format:

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

Example:

1,7,749,217,34,125,1,1,1
1,8,721,344,71,120,1,1,1
1,9,847,352,50,151,1,1,1
2,0,85,421,88,131,1,1,1

Field descriptions:

Field	Description
frame	Frame index, starting from 1
id	Player identity ID within the sequence
bb_left	Left coordinate of the bounding box
bb_top	Top coordinate of the bounding box
bb_width	Bounding-box width
bb_height	Bounding-box height
conf	Ground-truth confidence / validity flag
class	Object class label; players are the target class
visibility	Visibility value or flag, depending on the original annotation

Tracking Result Format

Many MOT evaluation tools expect tracker outputs in MOTChallenge result format:

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

For 2D tracking, the final world-coordinate fields are commonly filled with -1 if unused:

1,1,749,217,34,125,0.98,-1,-1,-1

Check the evaluation script or challenge instructions used for your experiment, because some SportsMOT conversion tools may expect a specific directory layout for TrackEval.

Usage

Download from Hugging Face

from huggingface_hub import snapshot_download

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

print(repo_dir)

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

Example: Read MOT Annotations with Python

from pathlib import Path
import pandas as pd

seq_dir = Path("SportsMOT/dataset/train/v_-6Os86HzwCs_c001")
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())
print(gt["id"].nunique(), "unique player tracks")

Example: Convert Bounding Boxes to xyxy

gt["x1"] = gt["bb_left"]
gt["y1"] = gt["bb_top"]
gt["x2"] = gt["bb_left"] + gt["bb_width"]
gt["y2"] = gt["bb_top"] + gt["bb_height"]

boxes_xyxy = gt[["frame", "id", "x1", "y1", "x2", "y2"]]
print(boxes_xyxy.head())

Example: Iterate Over Frames and Annotations

from pathlib import Path
import pandas as pd

seq_dir = Path("SportsMOT/dataset/train/v_-6Os86HzwCs_c001")
img_dir = seq_dir / "img1"
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"
    ],
)

for frame_path in sorted(img_dir.glob("*.jpg")):
    frame_idx = int(frame_path.stem)
    frame_boxes = gt[gt["frame"] == frame_idx]

    # Load frame_path with OpenCV, PIL, torchvision, etc.
    # Use frame_boxes for detection/tracking supervision.

Example: Read Split Files

from pathlib import Path

root = Path("SportsMOT")
train_sequences = (root / "splits_txt" / "train.txt").read_text().splitlines()
val_sequences = (root / "splits_txt" / "val.txt").read_text().splitlines()
test_sequences = (root / "splits_txt" / "test.txt").read_text().splitlines()

print(len(train_sequences), len(val_sequences), len(test_sequences))

Evaluation

SportsMOT can be evaluated with MOT-style metrics such as HOTA, MOTA, IDF1, AssA, DetA, ID switches, and fragmentation. The official repository includes a conversion script for TrackEval-style evaluation:

scripts/sportsmot_to_trackeval.py

For official benchmark reporting, use the evaluation protocol and challenge platform specified by the SportsMOT authors. For local experiments, keep these points in mind:

• Train and tune on the train and val splits only.
• Do not use test-set ground truth for model selection.
• Report the exact detector setup and whether detections are public, private, or generated by your own model.
• Ensure tracker outputs use the expected MOTChallenge-style frame indices and identity IDs.
• Keep video sequence boundaries separate; identity IDs are sequence-local.

Dataset Creation

Source Data

SportsMOT was built from professional sports videos covering basketball, football/soccer, and volleyball. The paper describes the videos as 720p, 25 FPS clips collected from sports sources such as NCAA, Premier League, Olympics, and NBA content. Clips were manually selected to avoid shot changes within a sequence.

Annotation Guidelines

The original annotation process follows player-centric MOT rules:

• Annotate the athlete’s visible limbs and torso.
• Exclude objects touching the athlete, such as balls.
• Predict the full bounding box under partial occlusion when part of the athlete remains visible.
• Skip an athlete when a large portion of the torso is outside the view.
• Keep a unique identity ID for each player throughout a clip.
• Remove extremely small boxes, such as boxes with width or height below 5 pixels.

Intended Use

SportsMOT is intended for research and benchmarking in:

• Sports multi-object tracking
• Player localization and identity association
• Motion and appearance modeling
• Sports video analytics
• Robust tracking under camera motion and occlusion

Out-of-Scope Use

This dataset should not be used for:

• Identifying private individuals outside the dataset context
• Surveillance or biometric identification
• Commercial use that violates the dataset license
• Redistribution that violates the original SportsMOT terms
• Claims of official benchmark results without following the official evaluation protocol

Limitations and Biases

SportsMOT is focused on professional sports videos and may not generalize to all sports, amateur footage, broadcast styles, camera viewpoints, resolutions, or frame rates. The target objects are players on the playground/court; spectators, referees, coaches, and balls are generally not tracking targets.

The dataset emphasizes player association rather than generic human tracking. Models trained only on SportsMOT may overfit to sports uniforms, field/court layouts, broadcast camera motion, and the three included sports categories.

License and Terms

SportsMOT is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0) according to the official repository.

The official repository also states that users should not redistribute the SportsMOT dataset without prior written permission. Please consult the original SportsMOT repository and challenge terms before redistributing, modifying, or using the dataset commercially.

Citation

If you use SportsMOT, please cite the original paper:

@inproceedings{cui2023sportsmot,
  title={SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes},
  author={Cui, Yutao and Zeng, Chenkai and Zhao, Xiaoyu and Yang, Yichun and Wu, Gangshan and Wang, Limin},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={9921--9931},
  year={2023}
}

Acknowledgements

SportsMOT was created by researchers from the Multimedia Computing Group at Nanjing University. Please refer to the original project page and paper for full author information, benchmark details, and official updates.