BillboardLamac Dataset

About

The BillboardLamac Dataset is a collection of videos, images, and metadata created to support research in computer vision tasks involving roadside advertisements and driver attention. The dataset was constructed from driving sessions using Tobii Glasses 3 eye-tracking technology, worn by 8 drivers along a predefined urban route in Bratislava, Slovakia.

The dataset supports two main tasks:

• Object Detection: Detecting and localizing billboards in urban driving scenes.
• Gaze-Based Classification: Categorizing driver attention toward detected billboards into four fixation classes.

Dataset Structure

The dataset contains several directories. A brief description of these directories can be found below:

• dataset for detection/: Part of the dataset used for training object detector models. It contains a total of 1213 images split into three subdirectories (train, val, test) with a ratio of 0.655/0.164/0.181. Each image has its respective billboard labels (if present). This part of the dataset is structured in a YOLO format.

• eye tracker/: Contains the original videos recorded using Tobii Glasses 3 for nine different rides: 0, 2, 3, 4, 5, 6a, 6b, 7, and 8. Each folder corresponds to a ride taken by a different driver along the same route. Route 6 was divided into two segments: 6a and 6b.

• fixation on detected ads/: Includes images from each ride where the driver's gaze intersected with a detected billboard, indicating visual fixation. Organized by ride number, consistent with the eye tracker/ folder.

• ids for classification/: Provides labeled data for billboard gaze classification:

  • images_ids/: Contains 145 unique billboards labeled respectively with their unique ids (image_id). These image_ids are distributed into four predefined gaze duration categories: long, medium, short, and none.

  • features.csv contains metadata and precomputed spatial features related to each classified gaze instance.

• objects/: Contains images of individual billboards:

  • a-objects/: billboards from the base (main) route.
  • b-objects/: billboards from extended routes (rides 4, 7, and 8).

• tracks/: Contains the ground_truth/ directory with text files containing annotated bounding boxes coordinates per frame for each ride. Each line in all files has the following format: frame_number, image_id, left, top, width, height. Each file (0.txt, 2.txt, etc.) corresponds to one ride.

Possible Use Cases

Researchers and developers may use this dataset for:

• Training object detection models focused solely on billboard detection.

• Exploring gaze-based visual attention and classifying driver focus using image-level data.

• Combining object detection with gaze behavior to assess visual saliency in real-world environments.

• Analyzing tracking data and spatial metadata for understanding attention patterns in urban scenes.

Notes

• The classification set only includes 145 labeled image IDs grouped by fixation duration.

• All images and videos originate from real driving sessions in a controlled experiment.

• Data resolution is high enough for billboard-level detection and detailed scene analysis.

• This dataset was created as part of an academic research project at Comenius University in Bratislava.

Citation

If you use this dataset in your work, please cite our research.