README.md

---
license: apache-2.0
---

# QuadTrack

<!-- Provide a quick summary of the dataset. -->

QuadTrack is a dataset designed for multi-object tracking (MOT) research, with a focus on panoramic and long-span scenarios. It provides challenging tracking sequences that include drastic appearance variations, prolonged occlusions, and wide field-of-view distortions, enabling the development and evaluation of robust MOT algorithms.

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->



- **Curated by:** [HNU CVPU]
- **Funded by [optional]:** [National Natural Science Foundation of China (No.62473139 and No.12174341), Zhejiang Provincial Natural Science Foundation of China (Grant No. LZ24F050003) and Shanghai SUPREMIND Technology Co. Ltd.]
- **Shared by [optional]:** [HNU CVPU]
- **License:** [CC BY-NC 4.0]

### Dataset Sources [optional]

<!-- Provide the basic links for the dataset. -->

- **Repository:** [https://github.com/xifen523/OmniTrack]
- **Paper:** [https://arxiv.org/abs/2503.04565]
- **Demo:** [https://www.youtube.com/watch?v=Q3mvzBtkkeU]

## Uses

<!-- Address questions around how the dataset is intended to be used. -->

### Direct Use

<!-- This section describes suitable use cases for the dataset. -->

QuadTrack is designed for multi-object tracking (MOT) research, particularly in panoramic.


## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->
The dataset is organized into two main splits: train and test.
```bash
QuadTrack/
├── train/                # Training set
│   ├── img1/             # Training images (video frames)
│   └── gt/               # Ground-truth annotations (bounding boxes, IDs, etc.)
│
└── test/                 # Test set
    └── img1/             # Test images (no ground-truth provided)
```



## Dataset Creation

### Curation Rationale

<!-- Motivation for the creation of this dataset. -->

QuadTrack was created to address the limitations of existing multi-object tracking (MOT) datasets, which often focus on narrow field-of-view scenarios and short-term associations. In contrast, panoramic and long-span tracking poses unique challenges such as:

+ Prolonged occlusions leading to identity switches.

+ Wide field-of-view distortions caused by panoramic cameras.

+ Dramatic appearance variations across long sequences.

The dataset aims to provide a benchmark for developing algorithms that achieve long-term identity stability and robust re-identification in real-world panoramic environments.

### Source Data

<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

Collection: The video sequences were captured using panoramic and wide-angle cameras in complex real-world environments (e.g., urban traffic, crowded public areas).

+ Annotation:

  + Bounding boxes and unique object IDs were assigned frame-by-frame.

  + Annotations follow the standard MOTChallenge format for compatibility.

+ Processing:

  + Frames were extracted at fixed intervals to balance temporal resolution and storage.

  + Quality checks ensured consistency in ID assignment across long occlusions.

  + Tools used: https://www.cvat.ai/
#### Who are the source data producers?

<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->

The source videos were collected and annotated by the QuadTrack research team.

+ Producers: Internal annotation team trained for MOT labeling tasks.

+ Demographics: Not applicable, as the dataset focuses on object trajectories rather than personal or sensitive identity information.

+ Note: No personally identifiable information (PII) is included. The dataset is curated strictly for research purposes.


## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

While QuadTrack provides challenging panoramic multi-object tracking scenarios, several limitations and risks should be noted:

+ Domain bias: The dataset primarily consists of panoramic and wide field-of-view sequences. Models trained on QuadTrack may not generalize well to conventional narrow-angle tracking datasets.

+ Scene diversity: Although collected across different environments, the dataset may not cover all possible real-world scenarios (e.g., extreme weather, night-time, or thermal imagery).

+ Annotation errors: Despite quality control, occasional inaccuracies in bounding boxes or identity switches may exist, especially under heavy occlusion.

+ Ethical risks: As a vision dataset, improper use in surveillance or privacy-intrusive applications could raise ethical concerns.

### Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.

## Citation [optional]

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**
```bibtex
@inproceedings{luo2025omnidirectional,
  title={Omnidirectional Multi-Object Tracking},
  author={Luo, Kai and Shi, Hao and Wu, Sheng and Teng, Fei and Duan, Mengfei and Huang, Chang and Wang, Yuhang and Wang, Kaiwei and Yang, Kailun},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={21959--21969},
  year={2025}
}
```

**APA:**

```mathematica
Luo, K., Shi, H., Wu, S., Teng, F., Duan, M., Huang, C., Wang, Y., Wang, K., & Yang, K. (2025). Omnidirectional multi-object tracking. *Proceedings of the Computer Vision and Pattern Recognition Conference*, 21959–21969.

```



## Dataset Card Authors [optional]

xifen527

## Dataset Card Contact

kailun.yang@hnu.edu.cn, luokai@hnu.edu.cn