dronefreak/LISA-Traffic-Lights

LISA Traffic Light Dataset

Unofficial redistribution of the LISA Traffic Light Dataset under the original CC BY-NC-SA 4.0 license.

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.

This repository exists for two purposes:

1. To provide a more reliable/durable download mirror of the original dataset.
2. To provide ready-to-use YOLO (Ultralytics) and COCO (RF-DETR / roboflow-style) export formats derived from the original CSV annotations, for teams benchmarking modern object detectors (YOLO family, RF-DETR, etc.) without having to write their own conversion pipeline.

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

---

Dataset Description

Reliable detection and recognition of traffic lights (TLs) is a key capability for autonomous vehicles and Advanced Driver Assistance Systems (ADAS). Prior to this dataset's release, traffic-light-recognition (TLR) research was evaluated almost exclusively on small, private datasets collected independently by each research group, making cross-paper comparison and progress tracking nearly impossible.

The LISA Traffic Light Dataset was collected to provide a common, public benchmark for TLR research. It consists of 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 camera view is used; the stereo capability is not exploited) at 1280×960 resolution with a 66° field of view, under both day and night conditions with varying light and weather.

Per the original authors' summary, the full official release totals 43,007 frames and 113,888 annotated traffic lights, split into:

• 13 daytime training clips and 5 nighttime training clips (dayTrain/, nightTrain/)
• 4 daytime and 2 nighttime continuous test sequences (daySequence1/2, nightSequence1/2), ~23 minutes 25 seconds of continuous driving

Two annotation styles are provided per frame, as CSV files alongside the video-derived JPEG frames:

• frameAnnotationsBOX.csv — bounding box around the entire traffic light housing (this is the box you generally want for detection training). Enlarged from the BULB annotation for any lit-area smaller than 4×4 px.
• frameAnnotationsBULB.csv — bounding box around only the lit bulb area.

Each CSV row has the form:

Filename;Annotation tag;Upper left corner X;Upper left corner Y;Lower right corner X;Lower right corner Y;Origin file;Origin frame number;Origin track;Origin track frame number

The Annotation tag (class) values found in frameAnnotationsBOX.csv are:

go, goLeft, goForward, stop, stopLeft, warning, warningLeft

---

Changes from the Official Release

The original dataset contents are preserved unmodified. All original images, videos-derived frames, and CSV annotation files are redistributed byte-for-byte as released by the LISA lab (including the sample-dayClip6 / sample-nightClip1 example subsets bundled with the original archive).

In addition to the untouched original data, this repository adds a data/converted/ directory containing derived, non-destructive exports for common training pipelines:

• Reorganized the BOX annotations into Ultralytics YOLO format (data/converted/yolo/box/) and, separately, the BULB annotations (data/converted/yolo/bulb/), each with images/, labels/, and a data.yaml.
• Generated a COCO-format export (data/converted/coco_rfdetr/) with train/valid/test directories and _annotations.coco.json files, suitable for RF-DETR and other COCO-style training pipelines.
• Defined explicit train / val / test splits at the clip level (not randomly per-frame) to avoid near-duplicate consecutive video frames leaking between splits:
  • train: 11 daytime + 4 nighttime training clips
  • val: held-out training clips (dayClip8, dayClip9, nightClip5)
  • test: the 4 official test sequences (daySequence1/2, nightSequence1/2)
• Deduplicated sample-dayClip6 / sample-nightClip1 out of the derived splits, since they are exact subsets of dayClip6 / nightClip1 already present in dayTrain/nightTrain — they are kept on disk (untouched) but excluded from the YOLO/COCO exports to avoid train/test leakage.

No original label values were changed, and no original samples were added or removed from the source archive. Class counts below were recomputed by us from the official BOX annotation CSVs and may differ slightly from the headline "43,007 frames / 113,888 boxes" summary above, due to the sample-folder deduplication described.

Apart from the additional data/converted/ exports and this dataset card, the repository is identical to the official release.

---

Dataset Structure

└── data/
    └── converted/                    # derived exports added by this repository
        ├── yolo/
        │   ├── box/                  # Ultralytics YOLO export (housing bounding boxes)
        │   │   ├── data.yaml
        │   │   ├── images/{train,val,test}/
        │   │   └── labels/{train,val,test}/
        │   └── bulb/                 # Ultralytics YOLO export (lit-bulb bounding boxes)
        │       ├── data.yaml
        │       ├── images/{train,val,test}/
        │       └── labels/{train,val,test}/
        └── coco_rfdetr/               # COCO-format export (from the BOX annotations)
            ├── train/_annotations.coco.json
            ├── valid/_annotations.coco.json
            └── test/_annotations.coco.json

Classes (BOX annotations, 7 classes)

id	class name
0	go
1	goForward
2	goLeft
3	stop
4	stopLeft
5	warning
6	warningLeft

Derived split sizes (BOX annotations)

Split	Images	Bounding boxes	Source clips
train	17,081	44,657	11 daytime + 4 nighttime training clips
val	3,454	7,169	2 daytime + 1 nighttime held-out training clips
test	22,481	57,649	4 official test sequences (day/night)

All images are 1280×960 JPEG, consistent across the entire dataset.

---

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/

---

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.

This repository only redistributes the original dataset under the same license, additionally packaging YOLO/COCO training-ready exports for convenience. 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 the YOLO/COCO exports in data/converted/ added by this repository) must be distributed under the same license.

This repository, including its derived exports, 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, which has significantly contributed to research in traffic light detection and recognition for autonomous driving and ADAS.