LISA Traffic Light Dataset (YOLO Box Export)

Unofficial redistribution of the LISA Traffic Light Dataset's "box" annotation style, exported in YOLO format, under the original CC BY-NC-SA 4.0 license.

Dataset Description

• Homepage: https://www.kaggle.com/datasets/mbornoe/lisa-traffic-light-dataset
• Paper: https://doi.org/10.1109/TITS.2015.2509509

Disclaimer

This repository is not an official release of the LISA Traffic Light Dataset.

The dataset was created by Morten Bornø Jensen, Mark Philip Philipsen, Andreas Møgelmose, Thomas B. Moeslund, and Mohan M. Trivedi at the Laboratory for Intelligent and Safe Automobiles (LISA), University of California, San Diego. The original authors retain all rights to the dataset. This repository does not claim ownership of any images, videos, or annotations, and takes no credit for the collection or annotation effort.

Three-hop provenance. This redistribution is narrower than a direct mirror of the original release. It is built from the box (traffic-light-housing) YOLO export inside dronefreak/LISA-Traffic-Lights, a companion project by the same author as this one, which itself converted the LISA lab's original CSV annotations into YOLO/COCO training formats with clip-level train/val/test splitting. This repository redistributes only the box-variant YOLO subset of that work, for use with DetectionBench, plus one configuration-file fix. If you want the bulb-annotation variant, the COCO/RF-DETR export, or the untouched original archive, see dronefreak/LISA-Traffic-Lights directly.

If you use this dataset, please respect the license terms below and cite the original papers, not this repository.

---

Dataset Overview

Reliable detection and recognition of traffic lights is a key capability for autonomous vehicles and Advanced Driver Assistance Systems (ADAS). The LISA Traffic Light Dataset was collected to provide a common, public benchmark for this task: continuous training and test video sequences recorded in San Diego, California (Pacific Beach and La Jolla), captured with a Point Grey Bumblebee XB3 stereo camera (only the left view is used) at 1280x960 resolution, under both day and night conditions with varying light and weather.

Two annotation styles exist upstream: BOX (bounding box around the entire traffic-light housing) and BULB (bounding box around only the lit bulb area). This repository redistributes only the BOX variant, already exported to YOLO format.

---

Changes from the Official Release

This is a narrower re-export of an already-derived work, not a direct mirror of the LISA lab's original release:

1. Original LISA release → dronefreak/LISA-Traffic-Lights (not performed by this repository)

The companion project converted the LISA lab's original per-frame CSV annotations (frameAnnotationsBOX.csv / frameAnnotationsBULB.csv) into Ultralytics YOLO and COCO formats, with explicit clip-level train/val/test splits (not per-frame random splits) to avoid near-duplicate consecutive video frames leaking across splits, and de-duplicated the bundled sample-dayClip6 / sample-nightClip1 example subsets against the full clips already present elsewhere in the data. See that repository for full details of this conversion.

2. dronefreak/LISA-Traffic-Lights (box/YOLO subset) → this repository

• Fixed a broken data.yaml: the source file's path: value is both relative and inconsistently cased, which does not resolve on a case-sensitive filesystem once relocated. A corrected data.yaml with an explicit, correctly-cased root path is provided here.
• Only the box-variant YOLO export was carried over; the bulb variant and COCO/RF-DETR export were not (see the companion repository for those).
• No images were added, removed, or modified. No labels were changed. No splits were changed.

Box counts below (44,657 / 7,169 / 57,649 for train/val/test) match the companion repository's own reported BOX YOLO export counts exactly, confirmed via DetectionBench's own YOLO→COCO conversion tooling.

---

Dataset Structure

dataset/
├── README.md
├── data.yaml
├── train/
│   ├── images/
│   └── labels/
├── val/
│   ├── images/
│   └── labels/
└── test/
    ├── images/
    └── labels/

where:

• images/ contains 1280x960 JPEG frames for each split.
• labels/ contains one YOLO-format .txt annotation file per image (class x_center y_center width height, normalized).
• data.yaml is the Ultralytics dataset configuration file (class names, split paths).
• Splits: train 17,082 images / 44,657 boxes · val 3,454 images / 7,169 boxes · test 22,481 images / 57,649 boxes (43,017 images total).

Classes (7)

---

Dataset Sources

Original Papers

Vision for Looking at Traffic Lights: Issues, Survey, and Perspectives Morten Bornø Jensen, Mark Philip Philipsen, Andreas Møgelmose, Thomas B. Moeslund, Mohan M. Trivedi IEEE Transactions on Intelligent Transportation Systems, 17(7), 1800-1815, 2016. DOI: 10.1109/TITS.2015.2509509

Traffic Light Detection: A Learning Algorithm and Evaluations on Challenging Dataset Mark Philip Philipsen, Morten Bornø Jensen, Andreas Møgelmose, Thomas B. Moeslund, Mohan M. Trivedi IEEE 18th International Conference on Intelligent Transportation Systems (ITSC), 2341-2345, 2015. DOI: 10.1109/ITSC.2015.7313470

Official Resources

• Original Kaggle listing: https://www.kaggle.com/datasets/mbornoe/lisa-traffic-light-dataset
• LISA Lab (UC San Diego): https://cvrr.ucsd.edu/

Companion Project (direct source of this export)

• dronefreak/LISA-Traffic-Lights on Hugging Face — the full original archive plus YOLO (box + bulb) and COCO/RF-DETR exports: https://huggingface.co/datasets/dronefreak/LISA-Traffic-Lights

---

Attribution

All credit for collecting, recording, and annotating this dataset belongs entirely to the original LISA lab authors — Morten Bornø Jensen, Mark Philip Philipsen, Andreas Møgelmose, Thomas B. Moeslund, and Mohan M. Trivedi.

The YOLO/COCO conversion, clip-level splitting, and deduplication this repository builds on was done in the dronefreak/LISA-Traffic-Lights companion project.

This repository only redistributes the box-variant YOLO subset of that work, with one configuration-file fix, for use with DetectionBench. It does not modify, reinterpret, or take credit for the underlying data or annotations.

If you use this dataset in your research, please cite the original publications below.

---

License

The original LISA Traffic Light Dataset is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license, as specified on the official Kaggle listing.

Accordingly:

• Attribution to the original authors is required.
• Commercial use is prohibited.
• Any derivative work (including this YOLO export, and the companion project it is built from) must be distributed under the same license.

This repository is distributed under the same CC BY-NC-SA 4.0 license.

---

Citation

If you use this dataset, please cite:

@article{jensen2016vision,
  title={Vision for looking at traffic lights: Issues, survey, and perspectives},
  author={Jensen, Morten Born{\o} and Philipsen, Mark Philip and M{\o}gelmose, Andreas and Moeslund, Thomas Baltzer and Trivedi, Mohan Manubhai},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={17},
  number={7},
  pages={1800--1815},
  year={2016},
  doi={10.1109/TITS.2015.2509509},
  publisher={IEEE}
}

@inproceedings{philipsen2015traffic,
  title={Traffic light detection: A learning algorithm and evaluations on challenging dataset},
  author={Philipsen, Mark Philip and Jensen, Morten Born{\o} and M{\o}gelmose, Andreas and Moeslund, Thomas B and Trivedi, Mohan M},
  booktitle={Intelligent Transportation Systems (ITSC), 2015 IEEE 18th International Conference on},
  pages={2341--2345},
  year={2015},
  organization={IEEE}
}

---

Acknowledgements

We sincerely thank the original LISA lab authors for collecting, annotating, and publicly releasing this valuable benchmark, and credit the dronefreak/LISA-Traffic-Lights companion project for the YOLO/COCO conversion work this repository builds on.