Saliency and scanpath models are developed almost exclusively on natural photographs, yet paintings are composed by artists to guide the viewer's eye and differ from photographs in texture, colour and abstraction. We introduce \avatt, a free-viewing eye-tracking dataset on \nImages{} paintings spanning sixteen art movements and five centuries, with \nScanpaths{} scanpaths and \nFixations{} fixations from \nParticipants{} participants (\spPerImageMin--\spPerImageMax{} observers per painting). For every painting we release individual scanpaths with fixation timing and eye--screen distance, fixation maps, saliency maps smoothed at one degree of visual angle, temporal saliency videos and rich art metadata. We benchmark classic and deep saliency models against human agreement, revealing substantial headroom for art-aware models, and rank 44 scanpath-similarity measures, including a new language-free family of image-encoder metrics built on ten CNN, ViT and self-supervised encoders. With controls that remove viewing biases, we show that faces are fixated \faceEnrich× more than expected and within half a second, that local structure draws the eye, that composition rules fitted from image content predict where people look, and that art style shapes how much viewers agree. Dataset, code and a Python package are available at: https://huggingface.co/datasets/Amine-CV/AVAtt.
Figures & tables
Dataset
Domain
Images
Tracking
Scanpaths
Sal. maps
Temporal
MIT1003 [ 26 ]
Natural
1,003
eye
✓
✓
–
CAT2000 [ 5 ]
20 categories
4,000
eye
✓
✓
–
SALICON [ 24 ]
Natural
20,000
mouse
✓
✓
–
Le Meur et al . [ 44 ]
Paintings (5 styles)
150
eye
–
✓
–
AVAtt (ours)
Paintings (16 movements)
283
eye
✓
✓
✓
Table 1: Comparison with public free-viewing datasets.
Figure 2: Paintings per movement group (left) and agreement between observers by movement (right).
Participants
152
Age (years)
23.5 ± 8.0 (median 20, range 18–62)
Gender
110 men (72.4%), 42 women (27.6%)
Art expertise
96.1% novice, 5 amateur, 1 expert
Origin
13 countries; France 85.5%
Vision
normal or corrected; Ishihara screening
Table 2: Participants (questionnaire data).
Figure 3: Frames of a temporal saliency video: attention accumulated up to t .
Figure 4: Fixation duration, saccade amplitude, eye–screen distance per fixation (shaded: 50–65 cm) and observers per painting.
Figure 5: Temporal dynamics over the first 5 s (mean ± 95% CI): fixation duration and saccade amplitude, inter-observer agreement, and convergence of the cumulative saliency map.
Predictor
AUC-J
sAUC
NSS
CC
SIM
KL ↓
Human (other half)
0.807
0.713
1.577
0.807
0.722
0.551
Single observer
0.737
0.654
1.084
0.548
0.498
4.206
DeepGaze IIE
0.771
0.620
1.217
0.695
0.638
0.589
TranSalNet-Res
0.771
0.649
1.205
0.680
0.658
0.533
UNISAL
0.776
0.654
1.270
0.695
0.659
0.600
GBVS
0.744
0.594
0.943
0.559
0.618
0.524
Table 3: Saliency benchmark on the paintings (mean over paintings, split-half). Bold: best non-human predictor.
Measure
Intra
Inter
AUC
rrb
% img.
q
FDISS
↑
0.1225
0.04939
0.796
1.00
99.6
<10−3
Mannan
↓
0.008298
0.01289
0.691
0.97
95.4
<10−3
EyeAnalysis
↓
0.03062
0.04701
0.687
0.97
94.7
<10−3
MultiMatch position
↑
0.8319
0.803
0.625
0.93
90.5
<10−3
ScanMatch
↓
100.4
103.6
0.593
0.73
83.4
<10−3
Hausdorff
↓
0.2949
0.328
0.589
0.83
80.9
<10−3
Table 4: Spatial scanpath measures, intra vs inter pairs. rrb : rank-biserial effect size; % img.: paintings where intra pairs are more similar; q : BH-corrected.
Encoder
Training
BERTScore
DTW
Mean-pool
VGG-16
CNN (supervised)
0.724
0.637
0.710
ResNet-50
CNN (supervised)
0.753
0.665
0.728
ConvNeXt-T
CNN (supervised)
0.656
0.542
0.620
ViT-B/16 original (Dosovitskiy et al.)
ViT (supervised)
0.692
0.613
0.678
CLIP ViT-B/16 (OpenAI)
ViT (language-supervised)
0.694
0.600
0.672
MAE ViT-B/16
Self-supervised
0.683
0.614
0.669
Table 5: Image-encoder scanpath measures: AUC of intra vs inter pairs (all q<10−3 ). Bold: best per column.
Figure 6: Image features at fixated vs baseline locations (AUC; * q<0.05 ).
Category
Paint.
Area%
Fix.%
Base%
Enrich.
@1s
q
face
93
5.1
20.1
7.2
4.97
57 / 17
<10−3
person
189
14.0
26.7
18.5
3.12
50 / 30
<10−3
vessel / vehicle
51
4.9
10.2
5.7
2.67
22 / 11
<10−3
animal
89
11.2
21.0
14.4
2.26
40 / 24
<10−3
food / fruit
13
3.2
7.2
5.1
1.94
22 / 16
<10−3
building
7
42.7
56.0
53.6
1.28
91 / 73
0.701
Table 6: Attention to object categories. Enrich.: observed / baseline share of fixations; @1s: % of viewers reaching the category within 1 s, observed / baseline.
Family
Var.
Top
Bottom
ρ
pρ
Rule of thirds
1
1.05
0.89
0.22
<10−3
Golden grid
1
1.04
0.88
0.23
<10−3
Centre
1
0.77
0.63
0.12
0.075
Diagonal
2
1.18
0.93
0.29
<10−3
Golden triangles
2
1.13
0.99
0.19
0.001
Dynamic symmetry
1
1.02
0.97
0.12
0.044
Table 7: Composition families. Var.: variants; Top / Bottom: fixation enrichment on the template (observed / baseline) for the best- and worst-fitting quarter of paintings; ρ : correlation between image-only fit and enrichment.
Figure 7: Variance in viewing explained by style, composition, content, low-level statistics and format (cross-validated R2 beyond exposure duration). Left: each block alone. Right: unique contribution (* permutation p<0.05 ).
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : Participants: age, gender, self-reported art expertise and country of origin.
Figure S2 : Colourfulness [ 16 ] and spatial information [ 23 ] of the paintings (painting area only). Mean colourfulness 36.3 ± 17.2, SI 59.8 ± 23.7; Pearson r=0.29 .
Figure S3 : Hue, saturation and value at fixated pixels (bars) vs all painting pixels (line).
Figure S4 : Saccade direction × amplitude (absolute and relative) and inter-observer congruency per painting.
Figure S5 : Fixation density over all paintings ( σ=1∘ ), with marginals.
Figure S6 : Maps from k observers vs five held-out observers (260 paintings, mean ± 95% CI).
Figure S7 : Saliency benchmark on the paintings (mean ± 95% CI over paintings).
Figure S8 : Spatial scanpath measures: intra- vs inter-image pairs.
Figure S9 : All 44 scanpath measures ranked by AUC (intra vs inter).
Figure S20 : Viewing measures by style (number of paintings in parentheses).
Figure S21 : Composition by style. Left: mean image-only fit (z-scored; * family differs across styles). Middle: fixation enrichment on the template. Right: within-style correlation between fit and fixation enrichment.
Figure S22 : Objects by style: share of paintings containing each category (left) and fixation enrichment over the baseline where present (right).
Figure S23 : Image-feature AUC (fixated vs baseline) by style.
University of Amsterdam, Amsterdam, The Netherlands · Amazon AGI, United States of America · College of Business and Economics, University of Johannesburg, South Africa