Pith. sign in

REVIEW 2 major objections 41 references

Learning to See Like Humans: Gaze-Aligned Cycling Safety Prediction

T0 review · 2 major / 0 minor · reviewed 2026-06-30 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Gaze-guided vision transformers match state-of-the-art safety rankings while aligning attention maps with human eye fixations

desk verdict The paper adds gaze supervision to a pairwise safety ranking model but the abstract supplies no numbers or details to check if it works. read the letter →

arxiv 2605.24040 v1 pith:KMSXUX7J submitted 2026-05-21 cs.CV

Luís Maria Perdigão , Miguel Costa , Carlos Santiago , Manuel Marques This is my paper
classification cs.CV
keywords eye-trackinggaze-guidedattentioncyclingsafetyperceptionvisiontransformerspairwiserankingurbananalyticsalignment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper develops a framework called EG-PCS that folds eye-tracking signals into a vision-transformer pipeline for ranking street images by perceived cycling safety. It keeps the existing pairwise comparison setup but adds supervision that pushes the model's attention toward the same regions humans fixate on when judging safety. Experiments show the guided models keep ranking performance comparable to leading methods while their attention maps match human gaze patterns more closely. A sympathetic reader would care because the work suggests that forcing AI to look where people look can improve both the reliability and the transparency of models used for urban safety decisions.

What carries the argument

The Eye-Tracking-Guided Perceived Cycling Safety (EG-PCS) framework, which adds eye-tracking supervision to the attention layers of vision transformers in a pairwise ranking setup for street-view safety judgments.

What would settle it

On a held-out collection of street images, the gaze-supervised model's attention maps show lower similarity to recorded human fixations than those of an unsupervised baseline, or its pairwise ranking accuracy drops below the current state-of-the-art on an external test set.

Watch

Extended reading notes

Core claim

By supervising the attention mechanism of vision transformers with eye-tracking signals inside a pairwise learning pipeline for perceived cycling safety, the EG-PCS framework produces attention maps that more accurately reflect human visual attention behavior while achieving similar ranking performance to state-of-the-art approaches.

Load-bearing premise

Eye-tracking signals collected from a limited set of viewers on a limited set of images provide a reliable supervisory signal that generalizes to how new viewers judge cycling safety in new images.

Editorial extensions

If this is right

  • Gaze-guided models produce attention maps closer to human fixation patterns than standard models.
  • Ranking accuracy on perceived cycling safety remains comparable to existing state-of-the-art methods.
  • The approach increases interpretability of model decisions in perception-based urban analytics.
  • Eye-tracking supervision can be added to pairwise learning pipelines without sacrificing predictive performance.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If human gaze patterns prove consistent across cities, gaze supervision could reduce reliance on city-specific labeled data for safety models.
  • The same supervision technique could transfer to related perception tasks such as rating walkability or perceived noise levels.
  • Larger and more diverse eye-tracking collections might further tighten the alignment between model attention and population-level safety judgments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 0 minor

Summary. The paper introduces the Eye-Tracking-Guided Perceived Cycling Safety (EG-PCS) framework that integrates gaze data into a vision transformer-based pairwise learning pipeline for assessing cycling safety from street-view images. By supervising the model's attention with eye-tracking signals, it aims to align learned attention maps with human fixation patterns. The experiments claim that gaze-guided models achieve similar ranking performance to state-of-the-art approaches while better reflecting human visual attention behavior.

Significance. If the results hold, this work could enhance interpretability in urban perception models by linking AI attention to human gaze patterns, potentially leading to more reliable safety predictions for cycling. The approach builds on pairwise comparisons by adding an attention supervision component, which is a positive step toward more human-aligned models in computer vision applications for urban analytics.

major comments (2)
  1. [Abstract] The abstract states that 'gaze-guided models achieve similar ranking performance compared to state-of-the-art approaches' and 'producing attention maps that more accurately reflect human visual attention behavior' but provides no quantitative results, baselines, statistical tests, dataset sizes, or details on gaze data collection and alignment. This makes the central claim impossible to evaluate without the full methods and results sections.
  2. [Data and Experiments] The assumption that eye-tracking signals from a limited set of viewers on limited images provide a generalizable supervisory signal is load-bearing for the claim of improved attention alignment transferring to the ranking task. No details are given on viewer count, demographic diversity, cross-validation on held-out images or new viewers, leaving the generalizability unverified.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive comments on the abstract and the generalizability of the eye-tracking supervisory signal. We address each point below and indicate revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] The abstract states that 'gaze-guided models achieve similar ranking performance compared to state-of-the-art approaches' and 'producing attention maps that more accurately reflect human visual attention behavior' but provides no quantitative results, baselines, statistical tests, dataset sizes, or details on gaze data collection and alignment. This makes the central claim impossible to evaluate without the full methods and results sections.

    Authors: We agree that the abstract would benefit from including key quantitative highlights to support the claims at a glance. In the revised manuscript, we have updated the abstract to incorporate specific ranking performance metrics, attention alignment improvements, dataset sizes, and brief details on gaze data collection and alignment methods. revision: yes

  2. Referee: [Data and Experiments] The assumption that eye-tracking signals from a limited set of viewers on limited images provide a generalizable supervisory signal is load-bearing for the claim of improved attention alignment transferring to the ranking task. No details are given on viewer count, demographic diversity, cross-validation on held-out images or new viewers, leaving the generalizability unverified.

    Authors: We acknowledge that more explicit details on the eye-tracking data are needed to substantiate generalizability. The revised manuscript expands the Data and Experiments section to specify viewer count, demographic information, dataset size, and cross-validation on held-out images. Performance on new viewers is noted as requiring additional data collection and is listed as a limitation for future work. revision: partial

standing simulated objections not resolved
  • Full verification of generalizability to entirely new viewers, which would require new eye-tracking experiments beyond the current manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; empirical outcome of gaze supervision on ranking task

full rationale

The abstract and described framework present an empirical pipeline: eye-tracking signals are collected separately and used as supervision to align attention maps in a vision-transformer pairwise ranking model. Performance is compared to external state-of-the-art baselines, and alignment quality is measured against held-out human fixations. No equations or claims reduce a prediction to a fitted parameter by construction, no uniqueness theorems are invoked via self-citation, and no ansatz is smuggled in. The central result (comparable ranking + improved attention match) is therefore falsifiable against independent data and does not collapse to the input by definition.

Assumptions & free parameters 0 free parameters · 1 assumptions · 1 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The central claim rests on the unverified assumption that gaze data constitutes an appropriate supervisory signal and that the reported performance gains are robust.

assumptions (1)
  • domain assumption Pairwise comparisons of street-view images provide a scalable way to learn subjective safety judgments.
    Stated as background from recent work in the abstract.
invented entities (1)
  • EG-PCS framework
    purpose: Integrates gaze data into pairwise learning pipeline based on vision transformers.
    New named method introduced in the abstract; no independent evidence provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to See Like Humans: Gaze-Aligned Cycling Safety Prediction." pith.science (2026). https://pith.science/paper/KMSXUX7J

@misc{pith2026260524040,
  author       = {Pith},
  title        = {Pith review of: Learning to See Like Humans: Gaze-Aligned Cycling Safety Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KMSXUX7J}},
  note         = {Machine review of arXiv:2605.24040}
}
read the original abstract

Cycling delivers significant public-health and environmental benefits, yet its uptake in cities is often limited by perceived safety. When street environments appear unsafe, individuals are less likely to cycle, making perception a key barrier to adoption. Recent work has shown that pairwise comparisons of street-view images provide a scalable way to learn subjective safety judgments. However, existing approaches do not explicitly model human visual attention, which plays a central role in how humans perceive safety. We propose an Eye-Tracking-Guided Perceived Cycling Safety framework (EG-PCS) that integrates gaze data into a pairwise learning pipeline based on vision transformers. By supervising the model's attention mechanism with eye-tracking signals, we encourage alignment between learned attention maps and human fixation patterns. Experiments show that gaze-guided models achieve similar ranking performance compared to state-of-the-art approaches while producing attention maps that more accurately reflect human visual attention behavior. Our results demonstrate that incorporating eye-tracking information enhances both predictive accuracy and interpretability in perception-based urban analytics.

Figures

Figures reproduced from arXiv: 2605.24040 by the authors.

Figure 1
Figure 1. Example of human vs. model attention on a street scene with a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Layout used in the eye-tracking experiment. Participants were asked [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. EG-PCS-Net architecture. Two input images, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison examples. From top to bottom: DINOv3, DeiT III, and CLIP. In each row (left-to-right): input image, eye-tracking gaze [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

  1. [1]

    Health benefits of cycling: a systematic review,

    P. Oja et al., “Health benefits of cycling: a systematic review,”Scan- dinavian Journal of Medicine & Science in Sports, vol. 21, no. 4, pp. 496–509, 2011

    work page 2011

  2. [2]

    Pucher and R

    J. Pucher and R. Buehler,City cycling. MIT press, 2012

    work page 2012

  3. [3]

    Bicycle commuting and facilities in major us cities: if you build them, commuters will use them,

    J. Dill and T. Carr, “Bicycle commuting and facilities in major us cities: if you build them, commuters will use them,”Transportation Research Record, vol. 1828, pp. 116–123, 2003

    work page 2003

  4. [4]

    Models of perceived cycling risk and route acceptability,

    J. Parkin, M. Wardman, and M. Page, “Models of perceived cycling risk and route acceptability,”Accident Analysis & Prevention, vol. 40, no. 3, pp. 1100–1107, 2008

    work page 2008

  5. [5]

    Route preferences among adults in the vancouver, bc, bicycle network,

    M. Winters and K. Teschke, “Route preferences among adults in the vancouver, bc, bicycle network,”American Journal of Health Promotion, vol. 25, no. 4, pp. 248–254, 2010

    work page 2010

  6. [6]

    Crash risk and subjective risk per- ception during urban cycling: Evidence for congruent and incongruent sources,

    R. von St ¨ulpnagel and J. Lucas, “Crash risk and subjective risk per- ception during urban cycling: Evidence for congruent and incongruent sources,”Accident Analysis and Prevention, vol. 142, 7 2020

    work page 2020

  7. [7]

    Assessing bikeability with street view imagery and computer vision,

    K. Ito and F. Biljecki, “Assessing bikeability with street view imagery and computer vision,”Transportation Research Part C: Emerging Tech- nologies, vol. 132, p. 103371, 2021

    work page 2021

  8. [8]

    Unpacking the perceived cycling safety of road environment using street view imagery and cycle accident data,

    Y . Ye, C. Zhong, and E. Suel, “Unpacking the perceived cycling safety of road environment using street view imagery and cycle accident data,” Accident Analysis and Prevention, vol. 205, 9 2024

    work page 2024

Show all 41 references
  1. [9]

    Streetscore: Predicting the perceived safety of one mil- lion streetscapes,

    N. Naik et al., “Streetscore: Predicting the perceived safety of one mil- lion streetscapes,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2014, pp. 779–785

    2014

  2. [10]

    Deep learning the city: Quantifying urban perception at a global scale,

    A. Dubey, N. Naik, J. Philipoom, R. Raskar, and C. A. Hidalgo, “Deep learning the city: Quantifying urban perception at a global scale,” in European Conference on Computer Vision (ECCV), 2016, pp. 196–212

    2016

  3. [11]

    Which cycling environment appears safer? learning cycling safety perceptions from pairwise image comparisons,

    M. Costa et al., “Which cycling environment appears safer? learning cycling safety perceptions from pairwise image comparisons,”IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 2, p. 1689–1700, Feb. 2025

    2025

  4. [12]

    Finding common image seman- tics for urban perceived safety based on pairwise comparisons,

    G. Costa, C. Soares, and M. Marques, “Finding common image seman- tics for urban perceived safety based on pairwise comparisons,” in2019 27th european signal processing conference. IEEE, 2019, pp. 1–5

    2019

  5. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy et al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  6. [14]

    Bridging the gap: a scoping review exploring the relationship between objective and subjective cycling safety,

    M. Costa, M. Marques, F. W. Siebert, C. Lima Azevedo, and F. Moura, “Bridging the gap: a scoping review exploring the relationship between objective and subjective cycling safety,”Transport Reviews, vol. 45, no. 6, pp. 993–1015, 2025

    2025

  7. [15]

    Cyclingnet: Detecting cycling near misses from video streams in complex urban scenes with deep learning,

    M. R. Ibrahim et al., “Cyclingnet: Detecting cycling near misses from video streams in complex urban scenes with deep learning,”IET Intelligent Transport Systems, vol. 15, no. 10, pp. 1331–1344, 2021

    2021

  8. [16]

    Cycling subjective experience: A conceptual framework and methods review,

    R. Zhang et al., “Cycling subjective experience: A conceptual framework and methods review,”Transportation Research Part F: Traffic Psychol- ogy and Behaviour, vol. 101, pp. 142–159, 2 2024

    2024

  9. [17]

    Using an immersive virtual reality bicycle simulator to evaluate hazard detection and anticipation of overt and covert traffic situations in young bicyclists,

    L. Zeuwts et al., “Using an immersive virtual reality bicycle simulator to evaluate hazard detection and anticipation of overt and covert traffic situations in young bicyclists,”Virtual Reality, vol. 27, pp. 1507–1527, 6 2023

    2023

  10. [18]

    FARSA: Fully automated roadway safety assessment,

    W. Song et al., “FARSA: Fully automated roadway safety assessment,” inIEEE Winter Conference on Applications of Computer Vision, 5 2018, pp. 521–529

    2018

  11. [19]

    Measuring heterogeneous perception of urban space with massive data and machine learning: An application to safety,

    T. Ram ´ırez et al., “Measuring heterogeneous perception of urban space with massive data and machine learning: An application to safety,” Landscape and Urban Planning, vol. 208, 4 2021

    2021

  12. [20]

    Using deep learning and google street view imagery to assess and improve cyclist safety in london,

    L. Rita et al., “Using deep learning and google street view imagery to assess and improve cyclist safety in london,”Sustainability, vol. 15, no. 13, p. 10270, 2023

    2023

  13. [21]

    Trueskill™: a bayesian skill rating system,

    R. Herbrich et al., “Trueskill™: a bayesian skill rating system,”Ad- vances in neural information processing systems, vol. 19, 2006

    2006

  14. [22]

    City-safe: Estimating urban safety perception,

    G. Costa, “City-safe: Estimating urban safety perception,” Master’s thesis, IST, 7 2019

    2019

  15. [23]

    A paired comparison approach for the analysis of sets of likert-scale responses,

    R. Dittrich, B. Francis, R. Hatzinger, and W. Katzenbeisser, “A paired comparison approach for the analysis of sets of likert-scale responses,” Statistical Modelling, vol. 7, no. 1, pp. 3–28, 2007

    2007

  16. [24]

    Assessing bikeability with street view imagery and computer vision,

    K. Ito and F. Biljecki, “Assessing bikeability with street view imagery and computer vision,”Transportation Research Part C: Emerging Tech- nologies, vol. 132, 11 2021

    2021

  17. [25]

    Tell me where to look: Guided attention inference network,

    K. Li et al., “Tell me where to look: Guided attention inference network,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018, pp. 9215–9223

    2018

  18. [26]

    Aegis: Human attention-based explainable guidance for intelligent vehicle systems,

    Z. Zhuang et al., “Aegis: Human attention-based explainable guidance for intelligent vehicle systems,” inConference on Human Factors in Computing Systems. Association for Computing Machinery, 4 2025

    2025

  19. [27]

    Gazevit: A gaze-guided hybrid attention vision transformer for cross-view matching of street-to-aerial images,

    Y . Hu et al., “Gazevit: A gaze-guided hybrid attention vision transformer for cross-view matching of street-to-aerial images,”Pattern Recognition Letters, vol. 191, pp. 80–88, 2025

    2025

  20. [28]

    Gaze-informed vision transformers: Predicting driving decisions under uncertainty,

    S. Koorathota et al., “Gaze-informed vision transformers: Predicting driving decisions under uncertainty,” in26th International Conference on Multimodal Interaction, 2024, pp. 184–194

    2024

  21. [29]

    Eye-gaze-guided vision transformer for rectifying shortcut learning,

    C. Ma, L. Zhao, Y . Chen, S. Wang, L. Guo, T. Zhang, D. Shen, X. Jiang, and T. Liu, “Eye-gaze-guided vision transformer for rectifying shortcut learning,”IEEE Transactions on Medical Imaging, vol. 42, no. 11, pp. 3384–3394, 2023

    2023

  22. [30]

    Gaze-guided vision transformer for chest x-ray image classification,

    Z. Chen, Z. Liu, and Y . Song, “Gaze-guided vision transformer for chest x-ray image classification,”Biomedical Signal Processing and Control, vol. 111, p. 108298, 2026

    2026

  23. [31]

    Ogama (open gaze and mouse analyzer): open-source software designed to an- alyze eye and mouse movements in slideshow study designs,

    A. V oßk ¨uhler, V . Nordmeier, L. Kuchinke, and A. M. Jacobs, “Ogama (open gaze and mouse analyzer): open-source software designed to an- alyze eye and mouse movements in slideshow study designs,”Behavior Research Methods, vol. 40, no. 4, pp. 1150–1162, 2008

    2008

  24. [32]

    Identifying fixations and saccades in eye-tracking protocols,

    D. D. Salvucci and J. H. Goldberg, “Identifying fixations and saccades in eye-tracking protocols,” inProceedings of the 2000 Symposium on Eye Tracking Research & Applications (ETRA). ACM, 2000, pp. 71–78

    2000

  25. [33]

    Methods for comparing scanpaths and saliency maps: Strengths and weaknesses,

    O. L. Meur and T. Baccino, “Methods for comparing scanpaths and saliency maps: Strengths and weaknesses,”Behavior Research Methods, vol. 45, pp. 251–266, 2013

    2013

  26. [34]

    What do different evaluation metrics tell us about saliency models?

    Z. Bylinskii, T. Judd, A. Oliva, A. Torralba, and F. Durand, “What do different evaluation metrics tell us about saliency models?” 4 2017. [Online]. Available: http://arxiv.org/abs/1604.03605

    2017 arXiv

  27. [35]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antigaet al., “Pytorch: An imperative style, high-performance deep learning library,” inAdvances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  28. [36]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” International Conference on Learning Representations (ICLR), 2019

    2019

  29. [37]

    Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour,

    P. Goyal et al., “Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  30. [38]

    Sgdr: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” inInternational Conference on Learning Representations (ICLR), 2017, arXiv:1608.03983

    2017 arXiv

  31. [39]

    Sim ´eoni et al., “Dinov3,”arXiv preprint arXiv:2508.10104, 2025

    O. Sim ´eoni et al., “Dinov3,”arXiv preprint arXiv:2508.10104, 2025

    2025 arXiv

  32. [40]

    Deit iii: Revenge of the vit,

    H. Touvron, M. Cord, and H. Jegou, “Deit iii: Revenge of the vit,”arXiv preprint arXiv:2204.07118, 2022

    2022

  33. [41]

    Learning transferable visual models from natural language supervision,

    A. Radford et al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PMLR, 2021, pp. 8748–8763

    2021

Pith tools

Reviewed June 30, 2026 · model on record in the stance chip above.