AVAtt: A Free-Viewing Eye-Tracking Dataset and Benchmark for Visual Attention on Paintings

1. Introduction

• Paintings differ from photographs in texture, palettes, abstraction and pictorial conventions, while artists deliberately organise subjects, lines and contrasts to guide viewing.Source: §1 · p. 2
• Public painting eye-tracking data remain scarce: the largest public collection covers 150 paintings in five styles with aggregated saliency maps but no individual scanpaths.Source: §1 · p. 2
• AVAtt releases 4,505 scanpaths on 283 paintings from sixteen movements, alongside fixation timing, viewing distance, maps, temporal videos and art metadata.Source: §1 · p. 2
• The paper benchmarks saliency and scanpath-similarity models and tests how faces, local structure, composition and style relate to attention.Source: §1 · p. 2

Eye-tracking datasets

• Free-viewing datasets such as MIT1003, CAT2000 [5] and SALICON drive saliency research, while CAT2000 includes an art category among twenty.Source: §2 · p. 2
• Le Meur et al. collected fixations on 150 paintings and found that saliency models transfer imperfectly to art; most other art-viewing studies use small, unreleased stimulus sets.Source: §2 · p. 2

Saliency models and scanpath prediction

• Saliency and scanpath work spans classic and deep models, sequence distances, FDISS, and language-model metrics; painting scanpath models have relied on adaptation from photographs.Source: §2 · p. 2

Composition

• Composition devices are intended to organise attention, but the link between automatically detected composition and gaze on paintings had not been tested at scale.Source: §2 · p. 2

Stimuli

• The released collection spans sixteen movement groups, the 15th–21st centuries and 81 identified artists, with artist, title, year, movement, genre and medium annotations.Source: §3 · p. 2

Participants and protocol

• Recordings came from 152 volunteers viewing paintings freely for 5.5 or 15 seconds with a 30 Hz eye tracker; the eye–screen distance averaged 56.8 ± 3.9 cm.Source: §3 · p. 3

Processing and quality control

• Quality control retained scanpaths spanning at least 4.5 seconds and paintings with at least ten valid scanpaths, yielding 283 paintings, 4,505 scanpaths and 121,426 fixations.Source: §3 · p. 3

Released data

• Each painting includes individual scanpaths, a binary fixation map, a saliency map smoothed at σ = 1° and a 60 fps temporal saliency video.Source: §3 · p. 3

Temporal dynamics and agreement

• 3.57 to 1.50: leave-one-out NSS agreement falls from the first to the fifth one-second window, while cumulative saliency-map correlation reaches 0.9 with the five-second map after 1.75 seconds.Source: §4 · p. 3
• Fixations lengthen and saccades shorten over the first five seconds, following the reported ambient-to-focal progression.Source: §4 · p. 3

Inter-observer agreement

• 0.60 ± 0.11: inter-observer congruency, measured as the share of fixations inside the others’ Otsu-thresholded map.Source: §4 · p. 3
• With three observers, AUC reaches 95% of its nine-observer value; CC and NSS reach that level with six observers.Source: §4 · p. 3

5.2. Scanpath similarity

• The scanpath test compares up to 45 same-painting pairs per painting against other-painting scanpaths mapped by relative position.Scanpaths are cut to five seconds, and the inter-image baseline retains realistic scanpath statistics.Source: §5.2 · p. 4

5.1. Saliency prediction

• UNISAL leads deep saliency models with NSS 1.27 and CC 0.695, while human agreement remains higher at CC 0.807 and NSS 1.58.The comparison uses split-half evaluation on painting areas; the Table 3 caption identifies the benchmark scope.Source: Table 3 · p. 4, §5.1 · p. 4

5.2. Scanpath similarity

• The encoder measures compare fixation-patch image embeddings using greedy patch-BERTScore, order-aware embedding DTW, or mean-pooled cosine across ten CNN, ViT, and self-supervised encoders.Each painting is encoded once, avoiding the thousands of generative calls per painting used by language-model-based metrics.Source: §5.2 · pp. 4–5
• FDISS is the most discriminative scanpath measure (AUC 0.796), followed by DINOv3 ViT-B/16 patch-BERTScore F1 (0.769).Intra-image pairs are more similar in 99.6% of paintings for FDISS; the encoder results compare intra- with inter-image pairs.Source: §5.2 · p. 5, Table 4 · p. 5, Table 5 · p. 5

6. What Draws Attention in Paintings

• Faces cover 5.1% of paintings containing them but receive 20.1% of fixations, versus 7.2% under the other-paintings baseline.Median first fixation on a face is 0.39 s versus 3.92 s for the baseline; local edge density and spatial information also attract attention most among tested image features.Source: §6 · p. 5, Table 6 · p. 6
• Image-only composition fits predict fixation enrichment on template points and lines across every family (Spearman ρ = 0.12–0.29), and the selected spiral variant attracts more fixations than the other seven.The best true-ratio spiral eye is reached 0.42 s earlier than baseline predicts; saccades also align with dominant painting lines.Source: §6 · p. 6, Table 7 · p. 6
• Across paintings, more faces correlate with lower saliency-map entropy (ρ = −0.31) and longer fixations (ρ = 0.38), while wide paintings correlate with more dispersed fixations (ρ = 0.78).Horizons correlate with more spread attention (ρ = 0.34), and diagonal lines with greater observer agreement (ρ = 0.16).Source: §6 · p. 6

6.1. Style, composition and content

• Viewing agreement and other measures differ across styles, and faces are over-fixated in every style by 3–10×.The composition-fit and fixation link holds within most styles, rather than merely reflecting between-style differences.Source: §6.1 · p. 6
• Unique descriptor contributions vary by viewing measure: format leads for dispersion, content for fixation duration, style for observer agreement, and composition for central tendency.Composition also adds to entropy prediction; its central-tendency contribution is 0.091, while style contributes 0.065 to observer agreement.Source: §6.1 · p. 6, Figure 7 · p. 7

7. Conclusion

• AVAtt provides individual scanpaths, saliency maps, temporal saliency videos, and art metadata for 283 paintings from sixteen movements.The paper also provides saliency and scanpath benchmarks and analyses of what guides attention in art.Source: §7 · p. 7
• The conclusion reports that paintings leave headroom for saliency models, while FDISS and the language-free encoder metrics best capture shared viewing.It also summarizes faces, local structure, composed salient content, and style as shaping where and how people look.Source: §7 · p. 7
• Future work includes viewers with art expertise and diverse viewer populations, higher-frequency recordings, and art-aware saliency, scanpath, and composition models trained on AVAtt.The dataset, code, and avatt package are made available.Source: §7 · p. 7

Supplementary Material

• Participant characteristics presented include age, gender, self-reported art expertise, and country of origin.The figure caption identifies these dimensions without reporting their values in the supplied passage.Source: Figure S1 · p. 10
• Participants were students, researchers, and staff recruited by convenience sampling without compensation.All participants gave informed consent, and the released data contain no personal information.Source: A. Participants and protocol · p. 10
• Colour vision was screened with three Ishihara plates; a few participants had mild red–green deficiencies without visible effects on gaze.The tracker was 50–70 cm from participants, and 90.4% of fixations occurred at 50–65 cm.Source: A. Participants and protocol · p. 10

• Mean colourfulness is 36.3 ± 17.2 and mean spatial information is 59.8 ± 23.7; their Pearson correlation is r = 0.29.Source: Figure S2 · p. 10

• Composition fit combines percentile-ranked point and curve fits, then retains the best template variant for each painting.The point fit measures salience-weighted proximity of focal subjects to template key points; the curve fit measures structure along template lines.Source: D. Content and composition · p. 10
• Figure S3 compares hue, saturation, and value at fixated pixels with values across all painting pixels.Source: Figure S3 · p. 11
• Figure S4 presents saccade direction and amplitude, in absolute and relative terms, alongside per-painting inter-observer congruency.Source: Figure S4 · p. 11

E. Style analysis

• The supplemental figures also present overall fixation density, held-out-observer maps, saliency-model benchmarks, and rankings of scanpath measures.Source: Figure S5 · p. 12, Figure S6 · p. 12, Figure S7 · p. 13
• Composition analyses show image-only detections and template variants, then plot fixation enrichment, attention over time, and saccade alignment with lines or a golden spiral.Source: Figure S10 · p. 15, Figure S11 · p. 16, Figure S12 · p. 16
• Other viewing analyses plot time to first fixation, attention to faces and object categories, feature effects over time, and correlations between painting properties and viewing behavior.Source: Figure S14 · p. 17, Figure S16 · p. 18, Figure S17 · p. 18
• Across styles, the figures compare viewing measures, composition fit and fixation enrichment, object prevalence and fixation enrichment, and feature AUC.Movement groups with fewer than five paintings are omitted; viewing measures use the first 5 s with exposure-duration adjustment, whereas feature AUCs use full viewing.Source: E. Style analysis · p. 11, Figure S20 · p. 20, Figure S21 · p. 20