pith. machine review for the scientific record. sign in

arxiv: 2605.03885 · v1 · submitted 2026-05-05 · 💻 cs.CV · cs.LG

Recognition: unknown

Raising the Ceiling: Better Empirical Fixation Densities for Saliency Benchmarking

Susmit Agrawal , Jannis Hollman , Matthias K\"ummerer
Authors on Pith no claims yet

Pith reviewed 2026-05-07 17:26 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords saliency benchmarkingfixation density estimationeye trackinginterobserver consistencykernel density estimationmixture modelscenter biasadaptive bandwidth
0
0 comments X

The pith

A per-image mixture model for fixation densities raises measured interobserver consistency by 5-15% over standard Gaussian kernels.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the decades-old method of estimating where people look from eye-tracking data systematically understates how much different observers agree on the same image locations. The proposed alternative fits a mixture of an adaptive-bandwidth kernel density estimate, center bias, uniform noise, and a state-of-the-art saliency model separately to each image using leave-one-subject-out cross-validation. This produces higher log-likelihood and AUC scores when predicting held-out human fixations, with the largest lifts occurring on the very images that current models fail on. The authors conclude that fixation density estimates are not static ground truth but can be improved with better statistical methods, which in turn changes how much headroom remains for saliency models.

Core claim

By modeling fixation locations as a mixture of an adaptive-bandwidth KDE based on Abramson's method, center bias and uniform components, and a state-of-the-art saliency model, with all parameters optimized per image via leave-one-subject-out cross-validation, the resulting densities achieve substantially higher agreement with held-out observers than fixed-bandwidth isotropic Gaussian KDE, yielding median per-image gains of 5-15% in log-likelihood and up to 2 percentage points in AUC, with improvements exceeding 25% on the most affected images.

What carries the argument

A per-image mixture model that combines Abramson's adaptive-bandwidth kernel density estimate, center bias, uniform distribution, and a state-of-the-art saliency model, with mixture weights and bandwidths optimized via leave-one-subject-out cross-validation to capture different types of interobserver consistency.

If this is right

  • Median per-image gains of 5-15% in log-likelihood and up to 2 percentage points in AUC when predicting held-out fixations.
  • Improvements exceed 25% precisely on the images most relevant to failure case analysis.
  • The new densities allow identification of remaining failure cases of state-of-the-art saliency models, indicating significant headroom for further model improvement.
  • Empirical fixation densities should be treated as evolving estimates that improve with better methodology rather than fixed ground truths.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Switching to these densities would likely re-rank which images count as current model failures and could shift priorities in future failure-case studies.
  • The same adaptive mixture approach could be applied to other point-pattern data, such as mouse clicks or trajectory traces, to obtain more reliable density estimates.
  • A controlled experiment that applies the method to new eye-tracking datasets without including any saliency model in the mixture would test how much of the reported gain depends on that component.
  • If widely adopted, benchmark conclusions about the sufficiency of existing saliency models would need to be revisited with the updated consistency ceiling.

Load-bearing premise

Including a state-of-the-art saliency model as one mixture component together with per-image parameter optimization via leave-one-subject-out cross-validation produces an unbiased estimate of human interobserver consistency rather than partially fitting the model to the benchmark data.

What would settle it

A side-by-side saliency benchmark comparison that shows whether the set of remaining model failure cases or the relative model rankings change when the new densities replace the standard Gaussian KDE densities, or a test that removes the saliency model component and checks whether the consistency gains largely disappear.

Figures

Figures reproduced from arXiv: 2605.03885 by Jannis Hollman, Matthias K\"ummerer, Susmit Agrawal.

Figure 1
Figure 1. Figure 1: (a) We propose a new model of empir￾ical fixation densities that shows that existing methods sometimes substantially underesti￾mated interobserver consistency. (b) This re￾stores headroom on saliency benchmarks for future progress (see Appendix G for details). Empirical fixation densities are not only used to report average benchmark performance; they in￾creasingly drive how we reason about models. First, … view at source ↗
Figure 2
Figure 2. Figure 2: Adaptive bandwidth KDE. Left: clas￾sic fixed-bandwidth Gaussian KDE. Right: Abramson adaptive bandwidth KDE, where each source fixation contributes with a dif￾ferent bandwidth determined by the pilot density estimate. Where fixations cluster, bandwidths shrink to preserve spatial detail; where fixations are sparse, bandwidths grow to remain smooth. In both cases, parameters are optimized for leave-one-subj… view at source ↗
Figure 3
Figure 3. Figure 3: Locally crossvalidated density. Pooling all fixations into a single KDE (left) overfits to individual fixation positions. Our locally weighted averaging of LOSO mod￾els (right) constructs a spatial density where each location is dominated by the fold that excluded the nearest observer, providing a more faithful density representation of interobserver consistency. are optimized jointly per image by maximizi… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of density visualization methods on three images with increasing peakedness. For each image, we visualize our empirical density with four different meth￾ods: Per-image heatmaps (first column) suggest strong patterns even in near-uniform distributions. Shared-scale heatmaps [7] (second column) enable comparison but ob￾scure detail. Quantile contours [33, 36, 38, 40] (third column) overemphasize s… view at source ↗
Figure 5
Figure 5. Figure 5: Interobserver Consistency Estimates: For four datasets, we compare estimates of inter-observer consistency (IOC) using different variants of our proposed model. We vary the KDE type (classic vs. Abramson adaptive bandwidth), the optimization procedure (global vs. per-image), and the mixture components (KDE only, KDE with uniform and center bias, and full model with additional DeepGaze MSDB component). For … view at source ↗
Figure 6
Figure 6. Figure 6: Images where LOSO prediction improves most from baseline to full model in terms of LL or AUC. structure, while AUC gains dominate on images with dispersed fixations, where the baseline overfits to sparse samples. This dichotomy has direct implications for sample-level analyses such as failure case studies or inverse benchmarking: the old density estimates were most inadequate precisely for the images where… view at source ↗
Figure 7
Figure 7. Figure 7: Per-component case studies. Starting with a classic Gaussian KDE with uniform and centerbias mixture components and parameters optimized on the full dataset, we (1) change to parameters optimized per image, (2) switch to Abramson adaptive bandwidth KDE, and (3) add the DeepGaze MSDB mixture component. For each step, we show the two images where performance improves most in terms of LL and AUC. See Appendix… view at source ↗
Figure 8
Figure 8. Figure 8: Cross-validation design for interobserver consistency estimation on MIT1003. (a) LOSO vs. LOFO cross-validation for our full model with different DeepGaze config￾urations. LOFO consistently yields higher consistency estimates, reflecting its access to subject pool information. The difference between the standard DeepGaze MSDB and a LOSO-trained variant is small, justifying the use of the standard model. (b… view at source ↗
Figure 9
Figure 9. Figure 9: Remaining SOTA model failures. We show the six images where DeepGaze MSDB [36] loses most performance compared to our improved empirical density, in terms of LL. Pixel-space information gain [37] reveals precisely where model predictions are most discrepant from the empirical density. We find clear failure categories such as paintings, context-dependent saliency, and semantic text importance. Figure 8b qua… view at source ↗
Figure 10
Figure 10. Figure 10: Extension to video: IOC estimates on LEDOV for different model variants and optimization granularities. Our full model reveals substantial headroom beyond what standard methodology suggested. Video saliency is far from solved. We extend our framework to video saliency on LEDOV (Fig￾ure 10). In the video setting, only ∼30ms of gaze data is available per subject and frame, making den￾sity estimation particu… view at source ↗
Figure 11
Figure 11. Figure 11: casestudy maximal improvement per model extension, Step 1: Starting with a classic Gaussian KDE with a bandwidth of one degree of visual angle, we change optimizing the bandwidth for leave-one-subject-out prediction and show the images where performance improves most in terms of LL and AUC. E Extended Per-Image Analysis For Model Components We provide extended per-image analyses for the case studies prese… view at source ↗
Figure 12
Figure 12. Figure 12: casestudy maximal improvement per model extension, Step 2: We now add uniform and centerbias mixture components and show the images where performance improves most in terms of LL and AUC. components takeover. On the other hand, it is largest (1.0) in cases with very strong and clear fixation clusters. For the weight of the uniform component, we see that it is largest (up to 0.15) for images with some very… view at source ↗
Figure 13
Figure 13. Figure 13: casestudy maximal improvement per model extension, Step 3: changing to parameters optimized per image. We show the images where performance improves most in terms of LL and AUC. H Extended Cross-Validation Design Analysis view at source ↗
Figure 14
Figure 14. Figure 14: casestudy maximal improvement per model extension, Step 4: Switching to Abramson adaptive bandwidth KDE. We show the images where performance improves most in terms of LL and AUC. LL: 0.49 bits/fix LL: 0.43 bits/fix LL: 0.41 bits/fix AUC: 4.55% AUC: 4.25% AUC: 4.18% Before 3.19 bits/fix 1.21 bits/fix 2.60 bits/fix 81.68% 79.49% 81.05% After 3.68 bits/fix 1.65 bits/fix 3.01 bits/fix 86.23% 83.74% 85.23% Ad… view at source ↗
Figure 15
Figure 15. Figure 15: casestudy maximal improvement per model extension, Step 5: Adding the DeepGaze MSDB mixture component. We show the images where performance im￾proves most in terms of LL and AUC. 26 view at source ↗
Figure 16
Figure 16. Figure 16: Extreme model parameters: Pilot bandwidth. We show the images where the pilot bandwidth is smallest or largest. param=0.01 param=0.01 param=0.01 param=27.16 param=25.44 param=25.07 Baseline model bandwidth=39.84 weights=0.82, 0.00, 0.17 LL: 1.37 bits bandwidth=39.84 weights=0.82, 0.00, 0.17 LL: 2.24 bits bandwidth=26.94 weights=0.88, 0.01, 0.11 LL: 2.46 bits bandwidth=43.66 weights=0.78, 0.01, 0.21 LL: 1.… view at source ↗
Figure 17
Figure 17. Figure 17: Extreme model parameters: Alpha. We show the images where the alpha parameter of the adaptive KDE is smallest or largest. param=0.00 param=0.00 param=0.00 param=0.93 param=0.92 param=0.91 Baseline model bandwidth=43.66 weights=0.78, 0.01, 0.21 LL: 1.48 bits bandwidth=39.84 weights=0.82, 0.00, 0.17 LL: 1.17 bits bandwidth=39.84 weights=0.82, 0.00, 0.17 LL: 1.49 bits bandwidth=26.94 weights=0.88, 0.01, 0.11… view at source ↗
Figure 18
Figure 18. Figure 18 view at source ↗
Figure 19
Figure 19. Figure 19: Extreme model parameters: Weight of the uniform component. We show the images where the weight is smallest or largest. param=0.00 param=0.00 param=0.00 param=0.93 param=0.75 param=0.63 Baseline model bandwidth=39.84 weights=0.82, 0.00, 0.17 LL: 2.16 bits bandwidth=39.84 weights=0.82, 0.00, 0.17 LL: 1.90 bits bandwidth=26.94 weights=0.88, 0.01, 0.11 LL: 2.66 bits bandwidth=43.66 weights=0.78, 0.01, 0.21 LL… view at source ↗
Figure 20
Figure 20. Figure 20: Extreme model parameters: Weight of the centerbias component. We show the images where the weight is smallest or largest. param=0.00 param=0.00 param=0.00 param=1.00 param=1.00 param=1.00 Baseline model bandwidth=43.66 weights=0.78, 0.01, 0.21 LL: 1.62 bits bandwidth=43.66 weights=0.78, 0.01, 0.21 LL: 2.29 bits bandwidth=26.94 weights=0.88, 0.01, 0.11 LL: 1.76 bits bandwidth=39.84 weights=0.82, 0.00, 0.17… view at source ↗
Figure 21
Figure 21. Figure 21 view at source ↗
Figure 22
Figure 22. Figure 22: Progress on the MIT300 dataset of the MIT/Tuebingen Saliency Benchmark in relation to our new estimates of interobserver consistency. See [34] for more details about model scoring. No DG LOSO DG All Subjects DG Deepgaze regularizer 2.1 2.2 2.3 2.4 2.5 2.6 2.7 IG [bit/fix] cv_type LOSO LOFO model upper crossval view at source ↗
Figure 23
Figure 23. Figure 23: Extended cross-validation design analysis. Solid lines show properly cross￾validated estimates; dashed lines show pooled (“upper”) estimates. The pooled esti￾mates overfit substantially in all configurations, with the largest overfitting without DeepGaze. The ordering between DeepGaze variants reverses between crossvalidated (solid) and pooled (dashed) evaluation, confirming that pooled estimates are domi… view at source ↗
read the original abstract

Empirical fixation densities, spatial distributions estimated from human eye-tracking data, are foundational to saliency benchmarking. They directly shape benchmark conclusions, leaderboard rankings, failure case analyses, and scientific claims about human visual behavior. Yet the standard estimation method, fixed-bandwidth isotropic Gaussian KDE, has gone essentially unchanged for decades. This matters now more than ever: as the field shifts toward sample-level evaluation (failure case analysis, inverse benchmarking, per-image model comparison), reliable per-image density estimates become critical. We propose a principled mixture model that combines an adaptive-bandwidth KDE based on Abramson's method, center bias and uniform components, and a state-of-the-art saliency model, to capture different spatial and semantic types of interobserver consistency, and optimize all parameters per image via leave-one-subject-out cross-validation. Our method yields substantially higher interobserver consistency estimates across multiple benchmarks, with median per-image gains of 5-15% in log-likelihood and up to 2 percentage points in AUC. For the most affected images -- precisely those most relevant to failure case analysis -- improvements exceed 25%. We leverage these improved estimates to identify and analyze remaining failure cases of state-of-the-art saliency models, demonstrating that significant headroom for model improvement remains. More broadly, our findings highlight that empirical fixation densities should not be treated as fixed ground truths but as evolving estimates that improve with better methodology.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript claims that standard fixed-bandwidth isotropic Gaussian KDE for empirical fixation densities is insufficient for modern saliency benchmarking, especially sample-level analysis. It proposes a four-component mixture (Abramson adaptive KDE + center bias + uniform + one external SOTA saliency model) whose weights and bandwidths are fit per image via LOOCV. This yields median per-image gains of 5-15% log-likelihood and up to 2 pp AUC in inter-observer consistency, with >25% gains on failure-case images, which the authors then use to argue that substantial headroom remains for SOTA saliency models.

Significance. If the estimator is unbiased, the work would be significant for the field: it would show that conventional densities systematically underestimate human consistency, revise failure-case analyses, and tighten the evaluation of model progress. The per-image optimization and explicit mixture of spatial/semantic components are methodologically interesting strengths. Credit is given for the focus on per-image metrics and the attempt to move beyond fixed KDE.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (mixture construction): the estimator includes a SOTA saliency model as an explicit component whose weight is chosen by per-image LOOCV on the same eye-tracking data used for benchmarking. This makes the resulting density a hybrid rather than a pure empirical human density. The central claims (higher consistency, >25% gains on failure cases, and 'significant headroom remains') are load-bearing on the assumption that this hybrid remains an unbiased human reference; the manuscript provides no quantitative test (e.g., weight distribution or ablation removing the saliency component) to show the assumption holds.
  2. [§4 and §5] §4 (experiments) and §5 (failure-case analysis): no ablation table isolates the contribution of the saliency-model component versus the adaptive KDE + bias terms alone. If the reported 5-15% log-likelihood and AUC gains are driven primarily by the external model, then the method does not demonstrate improved recovery of human density but rather injects model predictions, directly affecting the validity of the headroom conclusion.
minor comments (2)
  1. [Abstract] The abstract states 'multiple benchmarks' but does not name them; the methods section should list the exact datasets and splits used for the quantitative results.
  2. [§3] Notation for the mixture weights and the Abramson scaling factor should be defined once in §3 and used consistently; current presentation leaves the free-parameter count implicit.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive review. The concerns about the hybrid estimator and missing ablations are substantive and we will address them directly with additional analyses in the revision. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (mixture construction): the estimator includes a SOTA saliency model as an explicit component whose weight is chosen by per-image LOOCV on the same eye-tracking data used for benchmarking. This makes the resulting density a hybrid rather than a pure empirical human density. The central claims (higher consistency, >25% gains on failure cases, and 'significant headroom remains') are load-bearing on the assumption that this hybrid remains an unbiased human reference; the manuscript provides no quantitative test (e.g., weight distribution or ablation removing the saliency component) to show the assumption holds.

    Authors: We acknowledge that the inclusion of a SOTA saliency model renders the estimator hybrid. However, because all mixture weights (including the model weight) are optimized per image via LOOCV strictly on the held-out human fixations, the model component receives positive weight only when it improves the likelihood of the observed data; otherwise its weight is driven toward zero. This data-driven weighting preserves the estimator as an empirical reference conditioned on available priors. To directly test the concern, we will add (i) the distribution of learned component weights across all images and (ii) an ablation that removes the saliency-model term entirely. These additions will quantify how often and how much the model contributes and will support the claim that the resulting densities remain valid human references. revision: yes

  2. Referee: [§4 and §5] §4 (experiments) and §5 (failure-case analysis): no ablation table isolates the contribution of the saliency-model component versus the adaptive KDE + bias terms alone. If the reported 5-15% log-likelihood and AUC gains are driven primarily by the external model, then the method does not demonstrate improved recovery of human density but rather injects model predictions, directly affecting the validity of the headroom conclusion.

    Authors: We agree that an explicit ablation isolating the saliency-model term is required. In the revised manuscript we will insert a table (new Table X) that reports log-likelihood and AUC for four nested estimators on the same per-image LOOCV protocol: (1) fixed-bandwidth Gaussian baseline, (2) adaptive KDE + center bias + uniform, (3) adaptive KDE + center bias + uniform + saliency model, and (4) the full mixture. Preliminary internal results indicate that the adaptive KDE and bias terms already deliver the majority of the median 5-15 % gain, with the model term providing additional improvement primarily on semantically rich images; the full numbers will be reported. This ablation will allow readers to judge whether the headroom conclusion is driven by improved human-density recovery or by model injection. revision: yes

Circularity Check

1 steps flagged

Mixture model fits SOTA saliency component to human data, yielding hybrid densities used for benchmarking

specific steps
  1. fitted input called prediction [Abstract]
    "We propose a principled mixture model that combines an adaptive-bandwidth KDE based on Abramson's method, center bias and uniform components, and a state-of-the-art saliency model, to capture different spatial and semantic types of interobserver consistency, and optimize all parameters per image via leave-one-subject-out cross-validation. Our method yields substantially higher interobserver consistency estimates across multiple benchmarks, with median per-image gains of 5-15% in log-likelihood and up to 2 percentage points in AUC."

    The inter-observer consistency metric is computed from densities whose parameters are chosen by fitting a mixture that explicitly includes a SOTA saliency map as one component. The optimizer can therefore assign weight to the saliency map to improve the likelihood of held-out human fixations, making the reported consistency a hybrid human-plus-model quantity rather than a pure empirical human-human quantity. Subsequent use of these densities to benchmark saliency models (including models architecturally close to the included component) therefore evaluates models against an estimate that already incorporates model predictions by construction.

full rationale

The paper's method section (as described in the abstract) defines the empirical fixation density via a four-component mixture whose weights are optimized per-image on the eye-tracking data using LOOCV. Because one component is a fixed SOTA saliency map, non-zero weight on that component injects model predictions into the density estimate. The resulting densities are then treated as improved ground truth for measuring inter-observer consistency and for identifying model failure cases. This constitutes a fitted-input-called-prediction pattern: the density used for benchmarking is constructed in part from the very class of models being benchmarked. No equations reduce exactly to their inputs by algebraic identity, no uniqueness theorem is invoked via self-citation, and the adaptive KDE plus center-bias components remain independent. The circularity is therefore partial rather than total, warranting a score of 6.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that the four-component mixture plus per-image LOOCV recovers a better approximation to the true human fixation distribution than fixed-bandwidth KDE; this introduces several fitted parameters per image and the choice to include a saliency model as a density component.

free parameters (2)
  • mixture weights and bandwidths
    Optimized per image via leave-one-subject-out cross-validation; exact number and initialization not stated in abstract.
  • Abramson adaptive bandwidth scaling factor
    Inherited from Abramson's method but still tuned within the overall mixture optimization.
axioms (2)
  • domain assumption Fixation locations are independent given the image and the mixture parameters
    Implicit in any KDE or mixture density estimation of eye-tracking data.
  • ad hoc to paper A state-of-the-art saliency model can serve as a valid component of the human density without circular contamination
    Invoked when the mixture includes the saliency model; this is not a standard assumption in prior fixation-density literature.

pith-pipeline@v0.9.0 · 5554 in / 1721 out tokens · 48967 ms · 2026-05-07T17:26:34.562962+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 39 canonical work pages

  1. [1]

    URLhttps://doi.org/10.1214/aos/1176345338

    Abramson, I.S.: On Bandwidth Variation in Kernel Estimates-A Square Root Law. The Annals of Statistics10(4), 1217–1223 (Dec 1982).https://doi.org/10.1214/ aos/1176345986

    work page arXiv 1982

  2. [2]

    Vision Research46(18), 2824–2833 (Sep 2006).https://doi.org/10.1016/j.visres.2006.02.024

    Baddeley, R.J., Tatler, B.W.: High frequency edges (but not contrast) predict where we fixate: A Bayesian system identification analysis. Vision Research46(18), 2824–2833 (Sep 2006).https://doi.org/10.1016/j.visres.2006.02.024

    work page doi:10.1016/j.visres.2006.02.024 2006

  3. [3]

    Journal of Vision13(12), 1–1 (Oct 2013)

    Barthelmé, S., Trukenbrod, H., Engbert, R., Wichmann, F.: Modeling fixation locations using spatial point processes. Journal of Vision13(12), 1–1 (Oct 2013). https://doi.org/10.1167/13.12.1

    work page doi:10.1167/13.12.1 2013

  4. [4]

    Physica A: Statistical Mechanics and its Applications331(1), 207–218 (Jan 2004)

    Boccignone, G., Ferraro, M.: Modelling gaze shift as a constrained random walk. Physica A: Statistical Mechanics and its Applications331(1), 207–218 (Jan 2004). https://doi.org/10.1016/j.physa.2003.09.011

    work page doi:10.1016/j.physa.2003.09.011 2004

  5. [5]

    Future of Datasets

    Borji, A., Itti, L.: CAT2000: A Large Scale Fixation Dataset for Boosting Saliency Research. CVPR 2015 workshop on "Future of Datasets" (2015)

    2015

  6. [6]

    Technometrics19(2), 135–144 (May 1977).https://doi.org/10.1080/ 00401706.1977.10489521

    Breiman, L., Meisel, W., Purcell, E.: Variable Kernel Estimates of Multivariate Densities. Technometrics19(2), 135–144 (May 1977).https://doi.org/10.1080/ 00401706.1977.10489521

    work page arXiv 1977

  7. [7]

    Journal of Vision9(3), 5–5 (Mar 2009).https://doi.org/ 10.1167/9.3.5

    Bruce, N.D.B., Tsotsos, J.K.: Saliency, attention, and visual search: An information theoretic approach. Journal of Vision9(3), 5–5 (Mar 2009).https://doi.org/ 10.1167/9.3.5

    work page doi:10.1167/9.3.5 2009

  8. [8]

    1–1 (2018).https://doi.org/10.1109/ TPAMI.2018.2815601

    Bylinskii, Z., Judd, T., Oliva, A., Torralba, A., Durand, F.: What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1–1 (2018).https://doi.org/10.1109/ TPAMI.2018.2815601

    work page arXiv 2018

  9. [9]

    Bylinskii, Z., Recasens, A., Borji, A., Oliva, A., Torralba, A., Durand, F.: Where Should Saliency Models Look Next? In: Computer Vision – ECCV 2016. pp. 809–

    2016

  10. [10]

    Lecture Notes in Computer Science, Springer, Cham (Oct 2016).https:// doi.org/10.1007/978-3-319-46454-1_49

    work page doi:10.1007/978-3-319-46454-1_49 2016

  11. [11]

    Behavior Research Methods43(3), 864–878 (Sep 2011)

    Caldara, R., Miellet, S.: iMap: A novel method for statistical fixation mapping of eye movement data. Behavior Research Methods43(3), 864–878 (Sep 2011). https://doi.org/10.3758/s13428-011-0092-x

    work page doi:10.3758/s13428-011-0092-x 2011

  12. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, Y., Yang, Z., Chakraborty, S., Mondal, S., Ahn, S., Samaras, D., Hoai, M., Zelinsky, G.: Characterizing Target-Absent Human Attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5031–5040 (2022)

    2022

  13. [13]

    Vision Research102, 41–51 (Sep 2014).https://doi.org/10

    Clarke, A.D.F., Tatler, B.W.: Deriving an appropriate baseline for describing fix- ation behaviour. Vision Research102, 41–51 (Sep 2014).https://doi.org/10. 1016/j.visres.2014.06.016

    2014

  14. [14]

    Tutorials in Quantitative Methods for Psychology 1(1), 42–45 (Sep 2005).https://doi.org/10.20982/tqmp.01.1.p042 17

    Cousineau, D.: Confidence intervals in within-subject designs: A simpler solution to Loftus and Masson’s method. Tutorials in Quantitative Methods for Psychology 1(1), 42–45 (Sep 2005).https://doi.org/10.20982/tqmp.01.1.p042 17

    work page doi:10.20982/tqmp.01.1.p042 2005

  15. [15]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (Oct 2025)

    D’Agostino, F., Schwetlick, L., Bethge, M., Kuemmerer, M.: What Moves the Eyes: Doubling Mechanistic Model Performance Using Deep Networks to Discover and Test Cognitive Hypotheses. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (Oct 2025)

    2025

  16. [16]

    Journal of Vision10(10), 28–28 (Aug 2010).https://doi.org/10.1167/10.10.28

    Dorr, M., Martinetz, T., Gegenfurtner, K.R., Barth, E.: Variability of eye move- ments when viewing dynamic natural scenes. Journal of Vision10(10), 28–28 (Aug 2010).https://doi.org/10.1167/10.10.28

    work page doi:10.1167/10.10.28 2010

  17. [17]

    In: 2026 IEEE Winter Conference on Applications of Computer Vision (WACV) (2026)

    El-Jiz, P., Kümmerer, M., Tangemann, M., Bethge, M., Bartels, A., Bannert, M.: Isolating the Role of Temporal Information in Video Saliency: A Controlled Exper- imental Analysis. In: 2026 IEEE Winter Conference on Applications of Computer Vision (WACV) (2026)

    2026

  18. [18]

    Journal of Vision15(1), 14–14 (Jan 2015).https://doi.org/10.1167/15.1.14

    Engbert, R., Trukenbrod, H.A., Barthelmé, S., Wichmann, F.A.: Spatial statistics and attentional dynamics in scene viewing. Journal of Vision15(1), 14–14 (Jan 2015).https://doi.org/10.1167/15.1.14

    work page doi:10.1167/15.1.14 2015

  19. [19]

    completely blind

    Engelke, U., Liu, H., Wang, J., Le Callet, P., Heynderickx, I., Zepernick, H.J., Maeder, A.: Comparative Study of Fixation Density Maps. IEEE Transactions on Image Processing22(3), 1121–1133 (Mar 2013).https://doi.org/10.1109/TIP. 2012.2227767

    work page doi:10.1109/tip 2013

  20. [20]

    In: Che, W., Nabende, J., Shutova, E., Pilehvar, M.T

    Ghosh, A., Dziadzio, S., Prabhu, A., Udandarao, V., Albanie, S., Bethge, M.: ONEBench to Test Them All: Sample-Level Benchmarking Over Open-Ended Ca- pabilities. In: Che, W., Nabende, J., Shutova, E., Pilehvar, M.T. (eds.) Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers). pp. 32445–32481. As...

    work page doi:10.18653/v1/2025.acl-long.1560 2025

  21. [21]

    IEEE Transactions on Image Processing25(8), 3852–3861 (Aug 2016).https://doi.org/10.1109/TIP.2016

    Gide, M.S., Karam, L.J.: A Locally Weighted Fixation Density-Based Metric for Assessing the Quality of Visual Saliency Predictions. IEEE Transactions on Image Processing25(8), 3852–3861 (Aug 2016).https://doi.org/10.1109/TIP.2016. 2577498

    work page doi:10.1109/tip.2016 2016

  22. [22]

    Proceedings of the National Academy of Sciences p

    de Haas, B., Iakovidis, A.L., Schwarzkopf, D.S., Gegenfurtner, K.R.: Individual differences in visual salience vary along semantic dimensions. Proceedings of the National Academy of Sciences p. 201820553 (May 2019).https://doi.org/10. 1073/pnas.1820553116

    2019

  23. [23]

    Nicolás, The bar derived category of a curved dg algebra, Journal of Pure and Applied Algebra 212 (2008) 2633–2659

    Henderson, J.M.: Human gaze control during real-world scene perception. Trends in Cognitive Sciences7(11), 498–504 (Nov 2003).https://doi.org/10.1016/j. tics.2003.09.006

    work page doi:10.1016/j 2003

  24. [24]

    Scientific Reports9(1), 1–12 (Jan 2019).https://doi.org/10.1038/s41598-018- 37536-0

    Hoppe, D., Rothkopf, C.A.: Multi-step planning of eye movements in visual search. Scientific Reports9(1), 1–12 (Jan 2019).https://doi.org/10.1038/s41598-018- 37536-0

    work page doi:10.1038/s41598-018- 2019

  25. [25]

    In: Human Vision and Electronic Imaging IV

    Itti, L., Koch, C.: Comparison of feature combination strategies for saliency-based visual attention systems. In: Human Vision and Electronic Imaging IV. vol. 3644, pp. 473–483. International Society for Optics and Photonics (May 1999).https: //doi.org/10.1117/12.348467

    work page doi:10.1117/12.348467 1999

  26. [26]

    IEEE Transactions on pattern analysis and machine intelligence 20(11), 1254–1259 (1998)

    Itti, L., Koch, C., Niebur, E.: A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence 20(11), 1254–1259 (1998)

    1998

  27. [27]

    In: Proceedings of the European Conference on Computer Vision (ECCV)

    Jiang, L., Xu, M., Liu, T., Qiao, M., Wang, Z.: DeepVS: A Deep Learning Based Video Saliency Prediction Approach. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 602–617 (2018)

    2018

  28. [28]

    MIT Tech Report (Jan 2012) 18

    Judd, T., Durand, F., Torralba, A.: A Benchmark of Computational Models of Saliency to Predict Human Fixations. MIT Tech Report (Jan 2012) 18

    2012

  29. [29]

    In: Computer Vision, 2009 IEEE 12th International Conference On

    Judd, T., Ehinger, K., Durand, F., Torralba, A.: Learning to predict where humans look. In: Computer Vision, 2009 IEEE 12th International Conference On. pp. 2106–

    2009

  30. [30]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Kadner, F., Thomas, T., Hoppe, D., Rothkopf, C.A.: Improving Saliency Models’ Predictions of the Next Fixation With Humans’ Intrinsic Cost of Gaze Shifts. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2104–2114 (2023)

    2023

  31. [31]

    Trends in Cognitive Sciences4(4), 138–147 (Apr 2000).https://doi.org/10.1016/S1364-6613(00)01452-2

    Klein, R.M.: Inhibition of return. Trends in Cognitive Sciences4(4), 138–147 (Apr 2000).https://doi.org/10.1016/S1364-6613(00)01452-2

    work page doi:10.1016/s1364-6613(00)01452-2 2000

  32. [32]

    Koehler, K., Guo, F., Zhang, S., Eckstein, M.P.: What do saliency models predict? Journal of Vision14(3), 14–14 (Mar 2014).https://doi.org/10.1167/14.3.14

    work page doi:10.1167/14.3.14 2014

  33. [33]

    Kriegeskorte, N., Mur, M., Bandettini, P.A.: Representational similarity analysis - connectingthebranchesofsystemsneuroscience.FrontiersinSystemsNeuroscience 2(Nov 2008).https://doi.org/10.3389/neuro.06.004.2008

    work page doi:10.3389/neuro.06.004.2008 2008

  34. [34]

    Kümmerer, M.: Matthias-k/pysaliency (Feb 2026)

    2026

  35. [35]

    Annual Review of Vision Science9(1), 269–291 (Sep 2023).https://doi.org/10.1146/annurev-vision- 120822-072528

    Kümmerer, M., Bethge, M.: Predicting Visual Fixations. Annual Review of Vision Science9(1), 269–291 (Sep 2023).https://doi.org/10.1146/annurev-vision- 120822-072528

    work page doi:10.1146/annurev-vision- 2023

  36. [36]

    saliency.tuebingen.ai

    Kümmerer,M.,Bylinskii,Z.,Judd,T.,Borji,A.,Itti,L.,Durant,F.,Oliva,A.,Tor- ralba, A., Bethge, M.: MIT/Tuebingen Saliency Benchmark. saliency.tuebingen.ai

  37. [37]

    Kümmerer, M., Khanuja, H., Bethge, M.: Modeling Saliency Dataset Bias (May 2025).https://doi.org/10.48550/arXiv.2505.10169

    work page doi:10.48550/arxiv.2505.10169 2025

  38. [38]

    Proceedings of the National Academy of Sciences 112(52), 16054–16059 (Dec 2015).https://doi.org/10.1073/pnas.1510393112

    Kümmerer, M., Wallis, T.S.A., Bethge, M.: Information-theoretic model compar- ison unifies saliency metrics. Proceedings of the National Academy of Sciences 112(52), 16054–16059 (Dec 2015).https://doi.org/10.1073/pnas.1510393112

    work page doi:10.1073/pnas.1510393112 2015

  39. [39]

    In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y

    Kümmerer, M., Wallis, T.S.A., Bethge, M.: Saliency Benchmarking Made Easy: Separating Models, Maps and Metrics. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) Computer Vision – ECCV 2018. pp. 798–814. Lecture Notes in Computer Science, Springer International Publishing (2018)

    2018

  40. [40]

    In: The IEEE International Con- ference on Computer Vision (ICCV)

    Kümmerer, M., Wallis, T.S.A., Gatys, L.A., Bethge, M.: Understanding Low- and High-Level Contributions to Fixation Prediction. In: The IEEE International Con- ference on Computer Vision (ICCV). pp. 4789–4798 (2017)

    2017

  41. [41]

    In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision

    Linardos, A., Kümmerer, M., Press, O., Bethge, M.: DeepGaze IIE: Calibrated Pre- diction in and Out-of-Domain for State-of-the-Art Saliency Modeling. In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision. pp. 12919– 12928 (2021)

    2021

  42. [42]

    The Annals of Mathematical Statistics36(3), 1049–1051 (Jun 1965).https://doi.org/10.1214/aoms/1177700079

    Loftsgaarden, D.O., Quesenberry, C.P.: A Nonparametric Estimate of a Multivari- ate Density Function. The Annals of Mathematical Statistics36(3), 1049–1051 (Jun 1965).https://doi.org/10.1214/aoms/1177700079

    work page doi:10.1214/aoms/1177700079 1965

  43. [43]

    Behavior Research Methods45(1), 251–266 (Mar 2013)

    Meur, O.L., Baccino, T.: Methods for comparing scanpaths and saliency maps: Strengths and weaknesses. Behavior Research Methods45(1), 251–266 (Mar 2013). https://doi.org/10.3758/s13428-012-0226-9

    work page doi:10.3758/s13428-012-0226-9 2013

  44. [44]

    Tutorials in Quantitative Methods for Psychology4(2), 61–64 (Sep 2008).https://doi.org/10.20982/tqmp.04.2.p061

    Morey, R.D.: Confidence Intervals from Normalized Data: A correction to Cousineau (2005). Tutorials in Quantitative Methods for Psychology4(2), 61–64 (Sep 2008).https://doi.org/10.20982/tqmp.04.2.p061

    work page doi:10.20982/tqmp.04.2.p061 2005

  45. [45]

    In: 2013 IEEE International Conference on Computer Vision (ICCV)

    Riche, N., Duvinage, M., Mancas, M., Gosselin, B., Dutoit, T.: Saliency and Hu- man Fixations: State-of-the-Art and Study of Comparison Metrics. In: 2013 IEEE International Conference on Computer Vision (ICCV). pp. 1153–1160. IEEE (Dec 2013).https://doi.org/10.1109/ICCV.2013.147 19

    work page doi:10.1109/iccv.2013.147 2013

  46. [46]

    Journal of Vision17(13), 3–3 (Nov 2017).https://doi.org/10.1167/17.13.3

    Rothkegel, L.O.M., Trukenbrod, H.A., Schütt, H.H., Wichmann, F.A., Engbert, R.: Temporal evolution of the central fixation bias in scene viewing. Journal of Vision17(13), 3–3 (Nov 2017).https://doi.org/10.1167/17.13.3

    work page doi:10.1167/17.13.3 2017

  47. [47]

    Frontiers in Psychology15(May 2024)

    Schwetlick, L., Kümmerer, M., Bethge, M., Engbert, R.: Potsdam data set of eye movement on natural scenes (DAEMONS). Frontiers in Psychology15(May 2024). https://doi.org/10.3389/fpsyg.2024.1389609

    work page doi:10.3389/fpsyg.2024.1389609 2024

  48. [48]

    Communi- cations Biology3(1), 1–11 (Dec 2020).https://doi.org/10.1038/s42003-020- 01429-8

    Schwetlick, L., Rothkegel, L.O.M., Trukenbrod, H.A., Engbert, R.: Modeling the effects of perisaccadic attention on gaze statistics during scene viewing. Communi- cations Biology3(1), 1–11 (Dec 2020).https://doi.org/10.1038/s42003-020- 01429-8

    work page doi:10.1038/s42003-020- 2020

  49. [49]

    Journal of Vision11(1), 3 (Jan 2011)

    Smith, T.J., Henderson, J.M.: Looking back at Waldo: Oculomotor inhibition of return does not prevent return fixations. Journal of Vision11(1), 3 (Jan 2011). https://doi.org/10.1167/11.1.3

    work page doi:10.1167/11.1.3 2011

  50. [50]

    Journal of Vision7(14), 4–4 (Nov 2007).https://doi.org/10.1167/7.14.4

    Tatler, B.W.: The central fixation bias in scene viewing: Selecting an optimal view- ing position independently of motor biases and image feature distributions. Journal of Vision7(14), 4–4 (Nov 2007).https://doi.org/10.1167/7.14.4

    work page doi:10.1167/7.14.4 2007

  51. [51]

    Psychological Review124(3), 267–300 (2017).https: //doi.org/10.1037/rev0000054

    Tatler, B.W., Brockmole, J.R., Carpenter, R.H.S.: LATEST: A model of saccadic decisions in space and time. Psychological Review124(3), 267–300 (2017).https: //doi.org/10.1037/rev0000054

    work page doi:10.1037/rev0000054 2017

  52. [52]

    Tatler,B.W.,Vincent,B.T.:Systematictendenciesinsceneviewing.JournalofEye Movement Research2(2) (Dec 2008).https://doi.org/10.16910/jemr.2.2.5

    work page doi:10.16910/jemr.2.2.5 2008

  53. [53]

    Journal of Vision9(7), 4 (Jul 2009).https://doi.org/10.1167/9.7.4

    Tseng, P.H., Carmi, R., Cameron, I.G.M., Munoz, D.P., Itti, L.: Quantifying center bias of observers in free viewing of dynamic natural scenes. Journal of Vision9(7), 4 (Jul 2009).https://doi.org/10.1167/9.7.4

    work page doi:10.1167/9.7.4 2009

  54. [54]

    Visual Cognition17(6-7), 856–879 (Aug 2009)

    Vincent, B.T., Baddeley, R., Correani, A., Troscianko, T., Leonards, U.: Do we look at lights? Using mixture modelling to distinguish between low- and high-level factors in natural image viewing. Visual Cognition17(6-7), 856–879 (Aug 2009). https://doi.org/10.1080/13506280902916691

    work page doi:10.1080/13506280902916691 2009

  55. [55]

    In: Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition

    Volokitin, A., Gygli, M., Boix, X.: Predicting When Saliency Maps Are Accurate and Eye Fixations Consistent. In: Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition. pp. 544–552 (2016)

    2016

  56. [56]

    PLOS ONE6(9), e24038 (Sep 2011).https://doi.org/10

    Wilming, N., Betz, T., Kietzmann, T.C., König, P.: Measures and Limits of Models of Fixation Selection. PLOS ONE6(9), e24038 (Sep 2011).https://doi.org/10. 1371/journal.pone.0024038

    2011

  57. [57]

    Deriving regions of in- terest, coverage, and similarity using fixation maps

    Wooding, D.S.: Eye movements of large populations: II. Deriving regions of in- terest, coverage, and similarity using fixation maps. Behavior Research Methods, Instruments, & Computers34(4), 518–528 (Nov 2002).https://doi.org/10. 3758/BF03195481

    2002

  58. [58]

    A., and Elmar W., L

    Wu, Y., Konermann, H., Mededovic, E., Walter, P., Stegmaier, J.: Evaluating Cross-Subject and Cross-Device Consistency in Visual Fixation Prediction. In: 2025 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). pp. 1–7 (Jul 2025).https://doi.org/10.1109/ EMBC58623.2025.11252754

    work page arXiv 2025

  59. [59]

    Raising the Ceiling: Better Empirical Fixation Densities for Saliency Benchmarking

    Yang, Z., Mondal, S., Ahn, S., Zelinsky, G., Hoai, M., Samaras, D.: Predicting Human Attention using Computational Attention (Apr 2023).https://doi.org/ 10.48550/arXiv.2303.09383 20 Appendix for “Raising the Ceiling: Better Empirical Fixation Densities for Saliency Benchmarking” A Dataset Preprocessing In [36], the authors reported that the CAT2000 datase...

    work page doi:10.48550/arxiv.2303.09383 2023