🎨👁️ AVAtt — Art Visual Attention

Where do people look when they look at paintings?

Attention building up over the first five seconds of free viewing — 15+ observers per painting.

---

AVAtt is a free-viewing eye-tracking dataset on paintings. It spans 16 art movements, from Early Renaissance frescoes and Baroque masters to Ukiyo-e prints, Chinese ink, Cubism and abstract art. For every painting it provides individual scanpaths, fixation maps, saliency maps and temporal saliency videos, along with rich art metadata: title, artist, year, movement, genre and medium.

🖼️ 283 paintings

🎭 16 art movements

👩‍🎨 81 artists

👥 152 observers

〰️ 4,505 scanpaths

🎯 121,426 fixations

🎞️ 283 attention videos

✨ At a glance

Every painting comes with every fixation of every observer. Each fixation has its position, onset, duration and the viewer's distance from the screen. From these we derive a binary fixation map and a saliency map smoothed at one degree of visual angle (35.87 px), computed from the measured viewing geometry.

🎞️ Attention over time

Each painting has a 60 fps video of attention as it accumulates. Every fixation contributes the time already spent on it, so the frame at time t shows everything viewers have looked at up to t. In The Scream, the face is found within half a second and stays the focus, while exploration spreads to the sky and the railing. Rebuild any frame yourself: from avatt import maps maps.temporal_saliency_map(scanpaths, t_ms=1000)

🎭 Sixteen movements, five centuries

🚀 Quick start

With 🤗 datasets (images, maps, scanpaths and metadata):

from datasets import load_dataset

ds = load_dataset("Amine-CV/AVAtt", split="train")
row = ds[0]
row["title"], row["artist"], row["movement"]
row["image"]          # PIL, 1600×900 (painting centred on grey)
row["saliency_map"]   # PIL, σ = 1°
row["fixation_map"]   # PIL, binary
row["scanpaths"]      # [{observer, x, y, onset_ms, duration_ms, distance_mm}, ...]

With the avatt package (adds the videos, saliency metrics and map utilities):

pip install git+https://github.com/kmamine/AVAtt---Art-Visual-Attention.git

import avatt
from avatt import metrics

ds = avatt.load()                                    # downloads this repository once
painting = ds.image("avatt_0003", crop=True)         # The Scream, without the grey frame
scanpaths = ds.scanpaths("avatt_0003")               # one DataFrame per observer
score = metrics.nss(my_saliency_map, ds.fixation_map("avatt_0003"))

🏆 Benchmarks

Saliency prediction

Human split-half agreement vs. classic and deep saliency models, evaluated on the painting area.

Model	AUC-J ↑	sAUC ↑	NSS ↑	CC ↑	SIM ↑	KL ↓
Human (other half)	0.807	0.713	1.58	0.807	0.722	0.551
🥇 UNISAL	0.776	0.654	1.27	0.695	0.659	0.600
🥈 DeepGaze IIE	0.771	0.620	1.22	0.695	0.638	0.589
🥉 TranSalNet-Res	0.771	0.649	1.21	0.680	0.658	0.533
GBVS	0.744	0.594	0.94	0.559	0.618	0.524
Itti–Koch	0.724	0.621	0.84	0.489	0.592	0.561
Spectral residual	0.670	0.629	0.63	0.355	0.542	0.717
Centre bias	0.718	0.503	0.77	0.490	0.590	0.572

Open challenge: the best model is still well below human agreement. There is room for art-aware saliency models.

Scanpath similarity

How well do 44 measures recognise two viewers of the same painting? This includes FDISS, MultiMatch, ScanMatch, DTW and language-free image-encoder metrics built on DINOv3, I-JEPA, VICReg, CLIP and others.

Measure	AUC ↑	same-painting pairs win in
🥇 FDISS	0.796	99.6 % of paintings
🥈 DINOv3 patch-BERTScore	0.769	98.2 %
🥉 ResNet-50 / VICReg patch-BERTScore	0.753	98.9 %
Mannan	0.691	95.4 %
MultiMatch (position)	0.625	90.5 %

🔍 What draws the eye in paintings

• 🙂 Faces are fixated 5× more than expected and found within 0.4 s.
• 🧩 Local structure (edges, detail and contrast) attracts the eye more than colour.
• 📐 Composition guides gaze. The better a painting fits a template (thirds, golden spiral, diagonals…), the more viewers look at its points and lines.
• 🎭 Style shapes agreement. Chinese ink and Realism produce the most consistent viewing.

📂 Files

metadata.csv                         one row per painting
images/<id>.jpg                      1600×900 stimulus (painting centred on grey RGB 170)
scanpaths/<id>.jsonl                 {"observer": k, "fixations": [{"x", "y", "onset_ms", "duration_ms", "distance_mm"}, ...]}
fixation_maps/<id>.png               binary fixation map
saliency_maps/<id>.png               Gaussian-smoothed fixation map, σ = 1° of visual angle
temporal_saliency_maps/<id>.mp4      attention accumulated up to time t, 60 fps, 5 s
data/train-*.parquet                 everything above except the videos (viewer + load_dataset)
assets/                              figures of this card

Metadata columns

column	description
image_id	avatt_0001 … avatt_0283
title, artist, year	identification of the work
movement, movement_group	art movement and one of 16 movement groups
genre, medium	e.g. landscape / portrait; oil / woodblock print
attribution_confidence	high identified work · medium artist attributed · low movement only
copyright_status	public domain · in copyright · unknown
exposure_s	presentation time (5.5 s or 15 s)
n_observers, n_fixations	counts
painting_x0, painting_y0, painting_x1, painting_y1	the painting inside the 1600×900 canvas
duplicate_of, note, original_filename	curation notes

🔬 Collection

Eye tracker	Tobii X2-30 (30 Hz), 9-point calibration, I-VT fixation filter
Display	1920×1080, 531 mm wide; paintings on a 1600×900 grey canvas
Viewing distance	56.8 ± 3.9 cm, recorded for every fixation
Task	free viewing: "freely view and enjoy the paintings"
Presentation	5.5 s or 15 s per painting, separated by a grey screen
Observers	152 volunteers, 18–62 years, 13 countries, normal or corrected vision
Quality control	scanpaths ≥ 4.5 s, ≥ 10 observers per painting

⚖️ License

The gaze data, maps, videos and annotations are released under Permission to exploit licence with obligation to cite the paper, the data, and the preprints. Use them for anything.

The painting images are not covered by this licence. 248 works are in the public domain. The others, marked in copyright_status, are included solely to support non-commercial scientific research on visual attention. Rights holders can request removal through the Community tab.

📝 Citation

@inproceedings{kerkouri2024avatt,
  title     = {{AVAtt}: Art Visual Attention dataset for diverse painting styles},
  author    = {Kerkouri, Mohamed Amine and Tliba, Marouane and Chetouani, Aladine and Bruno, Alessandro},
  booktitle = {Proceedings of the 2024 Symposium on Eye Tracking Research and Applications (ETRA '24)},
  year      = {2024},
  publisher = {ACM},
  doi       = {10.1145/3649902.3655656}
}

_{Collected at the PRISME Laboratory, University of Orléans · funded by the TIC-ART project, Centre-Val de Loire Region, France}