AdGaze-3500

Overview

This dataset contains visual stimuli (images), features, and eye-tracking total gaze time data. It is designed for research in:

• Visual attention modeling
• Human perception
• Gaze time prediction

Structure

The dataset is organized into three main folders:

.
├── images/ # Raw images (ad and its context)
├── Features/ # Ad features
├── Gaze_Data/ # Total gaze times (on entire ad and ad elements) recorded by eye-tracking devices

1. data/

Contains all image stimuli shown to participants.

2. Features/

Contains ad features (e.g., ad locations, product category, etc.).

3. Gaze_Data/

Contains total gaze times, recorded by eye-tracking devices, on entire ad and each ad element.

---

Data Description

Images

• Format: .png
• Resolution: (1280, 1024)
• 3,531 images in total
• Naming convention: image_<id>.png

Features

• Format: .npy
• Description: Ad features, e.g. ad locations, ad sizes etc., more details please see the README.md in the Feature

Gaze Data

• Format: .npy
• Description: Total gaze times (in sec) on entire ad, brand element, text element and pictorial element.

---

Citation

@inproceedings{ye2026adgaze, title={AdGaze-3500: Evaluating Large Multimodal Models' Ability to Predict Human Attention to Ads}, author={Ye, Jianping and Wedel, Michel}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={4417--4427}, year={2026} }