Pith. sign in

REVIEW 3 major objections 4 minor 70 references

Accuracy does not buy human-likeness in monocular depth estimation

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 06:39 UTC pith:LNZBD65R

load-bearing objection The KITTI human-depth dataset is a real contribution and the inverse-U is worth arguing about, but the right-hand slope is not yet cleanly separated from a correlation artifact. the 3 major comments →

arxiv 2512.08163 v2 pith:LNZBD65R submitted 2025-12-09 cs.CV

Accuracy Does Not Guarantee Human-Likeness: Cross-Domain Human-Centered Benchmark in Monocular Depth Estimation

Yuki Kubota , Taiki Fukiage This is my paper
classification cs.CV
keywords monocular depth estimationhuman-AI alignmenthuman-like perceptionaccuracy-likeness trade-offKITTI benchmarkpartial correlationaffine decompositiondepth perception
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

A new human-centered benchmark asks people to judge absolute distances at fixed points in real street scenes and compares their error patterns with those of 69 deep-learning depth estimators. The paper's central claim is that metric accuracy and human-like behavior are not the same axis: as models become more accurate than humans, their error patterns diverge from human judgments, producing an inverse-U relationship in which models near human-level accuracy look most human. The same trade-off appears in the systematic components of error (scale, shift, shear, and residual) after an affine decomposition. The authors argue that physical accuracy and human-likeness are complementary, domain-sensitive dimensions of depth-estimation performance, so accuracy benchmarks alone cannot tell whether a model estimates depth the way people do.

Core claim

On KITTI, across 69 monocular depth estimators spanning architectures, training strategies, and datasets, the authors measured human-model similarity as the partial correlation between human and model depth estimates after removing the variance explained by physical depth, using scale-and-shift-aligned predictions. They found that most models show positive but substantially lower similarity to humans, and that human similarity does not rise monotonically with accuracy. In scale-recovered data the relationship is an inverse-U: among models less accurate than humans, improving accuracy is associated with higher human similarity (r=-0.56 between log-RMSE and similarity), while among models more

What carries the argument

The central measuring device is a partial correlation between human and DNN depth estimates after regressing out physical ground-truth depth, computed per image on 16 evaluation points with random half-splits of human observers; this isolates shared error patterns from accuracy. To dissect where the shared error lives, the paper applies an affine decomposition that fits each estimator's depth values to physical depth plus three geometric degrees of freedom — scale, shift, and horizontal and vertical shear — leaving a residual component. The inverse-U trade-off is then quantified by correlating log scale-shift-invariant RMSE with human similarity separately for models above and below human ac

Load-bearing premise

The trade-off rests on treating the correlation between the leftover errors of humans and models as a fair measure of human-likeness even though the most accurate models have very little leftover error left, which shrinks that correlation toward its noise floor.

What would settle it

Regress the residuals of the most accurate KITTI model onto human residuals, add a scaled copy of the human residual pattern back to the model's predictions so that its residual variance matches human residual variance, and recompute the partial correlation. If human similarity rises steeply in these variance-matched models, the right-hand side of the inverse-U is an artifact of residual shrinkage rather than a divergence of estimation strategy.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Accuracy benchmarks alone cannot certify perceptual alignment; a model can top the RMSE leaderboard while its error patterns are among the least human-like.
  • Models trained only on physical depth still pick up some human-like biases, but none reaches the reliability of human-human agreement.
  • The trade-off is not confined to one error type: it appears in scale, shift, shear, and residual components, so it reflects overall estimation strategy rather than a single artifact.
  • Human-likeness is domain-sensitive: rankings and error-space organization are only weakly preserved between outdoor KITTI and indoor NYU scenes, so a human-like model on one domain may not be human-like on another.
  • The benchmark supplies a reusable protocol for scoring future depth models on both metric accuracy and human similarity.

Where Pith is reading between the lines

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

  • If the inverse-U reflects a real strategy divergence, then training a high-accuracy model to mimic human error patterns should cost RMSE on standard benchmarks; this is directly testable by fine-tuning on human judgments.
  • The residual-variance caveat suggests part of the right-hand slope may be a measurement artifact: as a model's errors shrink, the residual correlation with human errors is attenuated toward a noise floor, so a variance-matched re-analysis is needed to confirm the trade-off.
  • The authors' KITTI-specific explanation — homogeneous road-sky geometry plus wide-angle distortion — predicts that on more diverse outdoor datasets the inverse-U should flatten or shift; collecting human judgments on such data would test that prediction.
  • If human-like representations indeed improve robustness, as argued in object recognition, human-similarity scores could become a cheap proxy for out-of-distribution robustness in depth models; the paper leaves that connection unestablished.

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

3 major / 4 minor

Summary. The paper proposes a human-centered evaluation framework for monocular depth estimation (MDE). The authors collected new human absolute-distance judgments at 16 points per image for KITTI scenes, compared 69 diverse MDE models against these judgments using partial correlations of residuals after controlling for physical depth, and further decomposed errors into affine components (scale, shift, shear, residual). They report an inverse-U relationship between metric accuracy (RMSE) and human similarity: models with accuracy around human level are most human-like, while both less accurate and more accurate models show lower similarity. The abstract and framing additionally claim a cross-domain benchmark including NYU Depth V2 data. The authors conclude that physical accuracy and human-likeness are complementary, domain-sensitive dimensions of MDE performance.

Significance. If the core trade-off claim is robust, this is a valuable contribution: it provides a large-scale outdoor human-depth dataset, a systematic comparison across 69 models, and a clear demonstration that accuracy benchmarks may not capture perceptual alignment. The paper is careful to fit per-image scale/shift alignment to physical ground truth rather than to human judgments, and the human-data collection is described in unusual detail. The release of data and analysis code would further strengthen the contribution. However, the central statistical evidence for the inverse-U is currently vulnerable to a measurement-artifact explanation, and the cross-domain framing is not fully supported by the manuscript as written.

major comments (3)
  1. [§3.3, §4.2, Table B1] The right-hand slope of the inverse-U may be a metric artifact. Human-DNN similarity is the partial correlation between residuals after regressing out physical depth. For the human-superior group, scale-recovered models have small residual variance, so any variance component not shared with human residuals (LiDAR noise, 16-point sampling, uint16 quantization, inference noise, human split-half noise) occupies a larger fraction of the residual and mechanically attenuates the correlation. The reported r=+0.69 (p<0.001) in the scale-recovered raw error, and the similar affine-component slopes, could therefore reflect shrinking residual variance rather than a genuine divergence of estimation strategies. The non-significant absolute-data right-hand slope (r=0.37, p=0.14) is consistent with this attenuation account. Please provide a noise-ceiling analysis or an amplitude-normalized similarity m
  2. [Abstract, §3.1, Appendix A.1] There is a direct numerical inconsistency: the abstract and Figure 1 state '652 KITTI and 654 NYU Depth V2 scenes,' while §3.1 and Appendix A.1 describe 328 KITTI evaluation images. Moreover, the NYU data collection is not described anywhere in this paper; it is only attributed to prior work [27]. As written, the cross-domain benchmark claim in the abstract is not supported by the methods. Please either include the full NYU data description and correct the scene counts, or revise the abstract and title claims to accurately reflect the KITTI-only scope of the new data and analyses.
  3. [§4.2 and Figure 1] The inverse-U is characterized by splitting models at the human RMSE and computing separate linear correlations within each group. This split is reasonable, but the paper should also report a continuous test of the non-monotonic relationship (e.g., quadratic regression or segmented regression with a breakpoint) to confirm that the decline on the right is statistically significant beyond a plateau. The current presentation relies on the reader accepting the human-RMSE boundary as the correct knot; a model-free test would strengthen the central claim.
minor comments (4)
  1. [Figure 1 and Figure 3 captions] The significance asterisks are mislabeled: '*:p<0.01, **:p<0.01, ***:p<0.001' should presumably be '*:p<0.05, **:p<0.01, ***:p<0.001'.
  2. [§3.3] The definition of 'partial correlation' would be clearer if it explicitly stated that both the human mean estimates and the DNN estimates are regressed on physical ground-truth depth before computing the correlation, and that for scale-recovered data these residuals are already orthogonal to depth by construction.
  3. [Table B1] The table reports many correlations without multiple-comparison correction. Given the large number of tests across components and conditions, a false-discovery-rate correction or at least a note about the risk of inflated significance would be appropriate.
  4. [Appendix B.1] The paragraph on residual error correctly notes that residual magnitudes are not comparable across models with different depth ranges. This caveat should also be reflected in the main-text discussion of affine residual components.

Circularity Check

0 steps flagged

No significant circularity: the inverse-U is an empirical measurement, not a quantity constructed from its inputs; self-citations to prior work are not load-bearing for the new KITTI result.

full rationale

The paper's central claim—that metric accuracy and human similarity exhibit an inverse-U relationship—is an observed correlation between two independently computed quantities. Human-DNN similarity is a split-half partial correlation between human and DNN residuals after regressing out physical depth (§3.3); accuracy is a scale-shift-invariant RMSE against physical ground truth (§3.3, §4.2). Neither quantity is fitted to the other, and no parameter is tuned to human judgments and then reported as a prediction. The affine decomposition fits the same linear model to humans and DNNs separately and then correlates the resulting coefficients (§3.4); this could produce a pattern but it is not a circular reduction because the coefficients are not derived from each other. The citation to the authors' prior work [27] supplies the NYU human-depth dataset and the affine-decomposition precedent, but the KITTI data collection, the 69-model evaluation, and all reported correlations are carried out in the present paper, so the self-citation is not load-bearing. The reader's concern about the right-hand slope (r=0.69) being attenuated by shrinking residual variance is a serious validity threat—the paper does not report a noise ceiling or an amplitude-normalized similarity—but it is a statistical confound, not circularity: a high-accuracy model with small residuals can in principle remain perfectly correlated with human residuals, so the observed divergence is not guaranteed by the definitions. The cross-domain abstract claim relies on NYU data cited from [27] rather than described here, which is a support/provenance gap, but not an equivalence of outputs to inputs.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The benchmark introduces no invented physical entities. Its free parameters are standard per-image alignment and decomposition fits plus a screening threshold; the load-bearing assumptions are psychophysical (crowdsourced absolute-distance judgments index human depth perception) and geometric (affine form captures the relevant error structure), inherited from the authors' prior work [27] and the pictorial-space literature.

free parameters (3)
  • Per-image scale/shift alignment (s, t) = per-image, per model (328 images × 69 models)
    Least-squares fit of each DNN output to LiDAR ground truth at the 16 evaluation points (Appendix A.3). Standard MDE normalization, but the alignment is fitted on the same 16 points used for the similarity score and it directly shrinks the residuals that drive the partial-correlation metric.
  • Affine decomposition coefficients (az, ax, ay, b) = per-image, per estimator (328 images × 4 coefficients)
    Four-parameter fit of estimated depth on ground-truth depth and pixel coordinates per image (§3.4). The residual error component is defined by this fit; the trade-off is then measured on these fitted quantities.
  • Participant reliability cutoff (1.5×IQR below pooled correlation median) = 1.5×IQR
    Hand-chosen screening threshold on participant-vs-group-median correlations (Appendix A.2); excludes 157 of 890 participants. Changing this threshold changes the human benchmark.
axioms (4)
  • domain assumption Crowdsourced absolute-distance judgments (meters, with scale bar and screen calibration) reliably measure human depth perception
    The entire benchmark treats mean crowd answers as the human reference. Validation is limited to split-half partial correlation (§3.3, Figure 2B); the text reports no comparison with established psychophysical depth estimates.
  • domain assumption Affine structure (scale, shift, two shears) captures the systematic component of both human and DNN depth errors
    Adopted from [27] and the pictorial-space literature [25,53] (§3.4). If relevant biases are non-affine, the component-wise trade-off analysis is incomplete.
  • standard math Partial correlation controlling for physical depth isolates 'error patterns' and is comparable across models with different residual magnitudes
    The regression-out computation is standard, but comparability is assumed without justification; the attenuation with shrinking residual variance is the paper's weakest unsupported premise (§3.3).
  • domain assumption LiDAR-measured depth in KITTI is the correct physical reference for both humans and models
    Errors for both humans and DNNs are defined against sensor ground truth (§3.1, §4.1), even though the paper's own premise is that human perception deviates systematically from physical depth.

pith-pipeline@v1.3.0-alltime-deepseek · 19980 in / 26940 out tokens · 599468 ms · 2026-08-04T06:39:30.922271+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Accuracy Does Not Guarantee Human-Likeness: Cross-Domain Human-Centered Benchmark in Monocular Depth Estimation." pith.science (2026). https://pith.science/paper/LNZBD65R

@misc{pith2026251208163,
  author       = {Pith},
  title        = {Pith review of: Accuracy Does Not Guarantee Human-Likeness: Cross-Domain Human-Centered Benchmark in Monocular Depth Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LNZBD65R}},
  note         = {Machine review of arXiv:2512.08163}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep neural networks (DNNs) are increasingly used as functional models of human vision, yet standard monocular depth estimation (MDE) benchmarks largely evaluate physical accuracy rather than behavioral alignment with humans. We introduce a cross-domain behavioral benchmark that aligns newly collected human absolute-distance judgments, physical ground truth, and model predictions at 16 locations in each of 652 KITTI and 654 NYU Depth V2 scenes. Human-model similarity was measured image-wise by partial correlation between scale-and-shift-aligned estimates while controlling for physical depth, and reference results were obtained for 69 diverse MDE models. Human error patterns were highly reliable, and most models showed positive but substantially lower similarity to humans. Human similarity, however, did not increase monotonically with metric accuracy: within each dataset, high-accuracy models converged toward similar residual error structures that remained distinct from humans, while both human-similarity scores and error-space organization were only weakly preserved across datasets. Thus, physical accuracy and human-like behavior are complementary, domain-sensitive dimensions of MDE performance. Our benchmark will provide a reusable basis for assessing future models along both dimensions.

Figures

Figures reproduced from arXiv: 2512.08163 by Taiki Fukiage, Yuki Kubota.

Figure 1
Figure 1. Figure 1: Overview of our research. We first constructed a human depth evaluation dataset based on an absolute depth estimation task. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of human and DNN error profiles for absolute (left) and scale-recovered (right) data. (A) RMSE of raw errors [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Scatter plots showing the relationship between model accuracy, quantified by scale-shift invariant RMSE, and human similarity [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of human and DNN depth estimation biases, illustrated with examples of characteristic horizontal ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

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

70 extracted references · 13 linked inside Pith

  1. [1]

    A study on the generality of neural network structures for monocu- lar depth estimation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2224–2238, 2023

    Jinwoo Bae, Kyumin Hwang, and Sunghoon Im. A study on the generality of neural network structures for monocu- lar depth estimation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2224–2238, 2023. 1, 2

    2023

  2. [2]

    The bas-relief ambiguity.International journal of computer vision, 35(1):33–44, 1999

    Peter N Belhumeur, David J Kriegman, and Alan L Yuille. The bas-relief ambiguity.International journal of computer vision, 35(1):33–44, 1999. 4

    1999

  3. [3]

    Adabins: Depth estimation using adaptive bins

    Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009–4018, 2021. 13, 17

    2021

  4. [4]

    Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023. 13, 14, 17

    Pith/arXiv arXiv 2023

  5. [5]

    Auto-rectify network for unsupervised indoor depth estimation.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 44(12): 9802–9813, 2021

    Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Tat-Jun Chin, Chunhua Shen, and Ian Reid. Auto-rectify network for unsupervised indoor depth estimation.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 44(12): 9802–9813, 2021. 13

    2021

  6. [6]

    Unsupervised scale-consistent depth learning from video.In- ternational Journal of Computer Vision, 129(9):2548–2564,

    Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Zhichao Li, Le Zhang, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth learning from video.In- ternational Journal of Computer Vision, 129(9):2548–2564,

  7. [7]

    MiDaS v3.1 – a model zoo for robust monocular relative depth estimation

    Reiner Birkl, Diana Wofk, and Matthias M¨uller. MiDaS v3.1 – a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 1, 13

    Pith/arXiv arXiv 2023

  8. [8]

    Single- image depth perception in the wild.Advances in Neural In- formation Processing Systems, 29, 2016

    Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single- image depth perception in the wild.Advances in Neural In- formation Processing Systems, 29, 2016. 1, 3

    2016

  9. [9]

    Scaling Vision Transformers to 22 billion pa- rameters

    Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdul- mohsin, et al. Scaling Vision Transformers to 22 billion pa- rameters. InInternational Conference on Machine Learning, pages 7480–7512, 2023. 1, 7

    2023

  10. [10]

    Omnidata: A scalable pipeline for making multi- task mid-level vision datasets from 3d scans

    Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi- task mid-level vision datasets from 3d scans. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 10786–10796, 2021. 13

    2021

  11. [11]

    Depth map prediction from a single image using a multi-scale deep net- work.Advances in Neural Information Processing Systems, 27, 2014

    David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep net- work.Advances in Neural Information Processing Systems, 27, 2014. 1, 2, 13

    2014

  12. [12]

    Adversar- ial robustness as a prior for learned representations.arXiv preprint arXiv:1906.00945, 2019

    Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Adversar- ial robustness as a prior for learned representations.arXiv preprint arXiv:1906.00945, 2019. 1

    Pith/arXiv arXiv 1906

  13. [13]

    Deeper into self-supervised monocular indoor depth estimation

    Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 13

    Pith/arXiv arXiv 2023

  14. [14]

    ImageNet-trained CNNs are biased towards texture; increas- ing shape bias improves accuracy and robustness

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. ImageNet-trained CNNs are biased towards texture; increas- ing shape bias improves accuracy and robustness. InInter- national Conference on Learning Representations, 2018. 1

    2018

  15. [15]

    General- isation in humans and deep neural networks.Advances in Neural Information Processing Systems, 31, 2018

    Robert Geirhos, Carlos RM Temme, Jonas Rauber, Heiko H Sch¨utt, Matthias Bethge, and Felix A Wichmann. General- isation in humans and deep neural networks.Advances in Neural Information Processing Systems, 31, 2018. 1

    2018

  16. [16]

    3D packing for self-supervised monocular depth estimation

    Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raven- tos, and Adrien Gaidon. 3D packing for self-supervised monocular depth estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2485–2494, 2020. 4, 13, 14

    2020

  17. [17]

    Towards zero-shot scale-aware monocu- lar depth estimation

    Vitor Guizilini, Igor Vasiljevic, Dian Chen, Rares , Ambrus,, and Adrien Gaidon. Towards zero-shot scale-aware monocu- lar depth estimation. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9233–9243,

  18. [18]

    Revealing the multidimensional mental rep- resentations of natural objects underlying human similarity judgements.Nature Human Behaviour, 4(11):1173–1185,

    Martin N Hebart, Charles Y Zheng, Francisco Pereira, and Chris I Baker. Revealing the multidimensional mental rep- resentations of natural objects underlying human similarity judgements.Nature Human Behaviour, 4(11):1173–1185,

  19. [19]

    Do Vision Transformers see like humans? Evaluating their perceptual alignment.arXiv preprint arXiv:2508.09850, 2025

    Pablo Hern ´andez-C´amara, Jose Manuel Ja ´en-Lorites, Jorge Vila-Tom´as, Valero Laparra, and Jesus Malo. Do Vision Transformers see like humans? Evaluating their perceptual alignment.arXiv preprint arXiv:2508.09850, 2025. 1, 7

    Pith/arXiv arXiv 2025

  20. [20]

    Im- mediate generalisation in humans but a generalisation lag in deep neural networks—evidence for representational diver- gence?arXiv preprint arXiv:2402.09303, 2024

    Lukas S Huber, Fred W Mast, and Felix A Wichmann. Im- mediate generalisation in humans but a generalisation lag in deep neural networks—evidence for representational diver- gence?arXiv preprint arXiv:2402.09303, 2024. 1

    Pith/arXiv arXiv 2024

  21. [21]

    A new cue to figure-ground coding: Top-bottom polarity.Vision Research, 44(24):2779–2791, 2004

    Johan Hulleman and Glyn W Humphreys. A new cue to figure-ground coding: Top-bottom polarity.Vision Research, 44(24):2779–2791, 2004. 14

    2004

  22. [22]

    Repurpos- ing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met- zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos- ing diffusion-based image generators for monocular depth estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492– 9502, 2024. 1, 12, 13

    2024

  23. [23]

    Global-local path net- works for monocular depth estimation with vertical cutdepth

    Doyeon Kim, Woonghyun Ka, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, and Junmo Kim. Global-local path net- works for monocular depth estimation with vertical cutdepth. arXiv preprint arXiv:2201.07436, 2022. 13, 17 9

    Pith/arXiv arXiv 2022

  24. [24]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything.arXiv:2304.02643, 2023. 12

    Pith/arXiv arXiv 2023

  25. [25]

    Ambiguity and the ‘mental eye’ in pictorial relief.Perception, 30(4):431–448, 2001

    Jan J Koenderink, Andrea J van Doorn, Astrid M L Kap- pers, and James T Todd. Ambiguity and the ‘mental eye’ in pictorial relief.Perception, 30(4):431–448, 2001. PMID: 11383191. 3, 4

    2001

  26. [26]

    Distance perception as a function of photographic area of view.Perception & Psychophysics, 45(5):459–466, 1989

    Robert N Kraft and Jeffrey S Green. Distance perception as a function of photographic area of view.Perception & Psychophysics, 45(5):459–466, 1989. 8, 14

    1989

  27. [27]

    Human-like monocular depth biases in deep neural networks.PLOS Computational Biology, 21(8):e1013020, 2025

    Yuki Kubota and Taiki Fukiage. Human-like monocular depth biases in deep neural networks.PLOS Computational Biology, 21(8):e1013020, 2025. 1, 2, 3, 4, 8

    2025

  28. [28]

    Motion-driven enhancement of a lower region cue in depth perception.Journal of Illusion, 3, 2022

    Yuki Kubota, Ryota Mima, Takahiro Kawabe, Taiki Fukiage, and Masahiko Inami. Motion-driven enhancement of a lower region cue in depth perception.Journal of Illusion, 3, 2022. 14

    2022

  29. [29]

    Do better ImageNet classifiers assess perceptual sim- ilarity better?arXiv preprint arXiv:2203.04946, 2022

    Manoj Kumar, Neil Houlsby, Nal Kalchbrenner, and Ekin D Cubuk. Do better ImageNet classifiers assess perceptual sim- ilarity better?arXiv preprint arXiv:2203.04946, 2022. 1, 7

    Pith/arXiv arXiv 2022

  30. [30]

    Measurement and modeling of depth cue combination: in defense of weak fusion.Vision Re- search, 35(3):389–412, 1995

    Michael S Landy, Laurence T Maloney, Elizabeth B John- ston, and Mark Young. Measurement and modeling of depth cue combination: in defense of weak fusion.Vision Re- search, 35(3):389–412, 1995. 3

    1995

  31. [31]

    EVP: Enhanced visual perception us- ing inverse multi-attentive feature refinement and regularized image-text alignment.arXiv preprint arXiv:2312.08548,

    Mykola Lavreniuk, Shariq Farooq Bhat, Matthias M ¨uller, and Peter Wonka. EVP: Enhanced visual perception us- ing inverse multi-attentive feature refinement and regularized image-text alignment.arXiv preprint arXiv:2312.08548,

    Pith/arXiv arXiv

  32. [32]

    Instance-wise occlusion and depth orders in natural scenes

    Hyunmin Lee and Jaesik Park. Instance-wise occlusion and depth orders in natural scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21210–21221, 2022. 1, 3

    2022

  33. [33]

    From big to small: Multi-scale local planar guidance for monocular depth estimation.arXiv preprint arXiv:1907.10326, 2019

    Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation.arXiv preprint arXiv:1907.10326, 2019. 13, 17

    Pith/arXiv arXiv 1907

  34. [34]

    Structdepth: Leveraging the structural regularities for self-supervised indoor depth estimation

    Boying Li, Yuan Huang, Zeyu Liu, Danping Zou, and Wenx- ian Yu. Structdepth: Leveraging the structural regularities for self-supervised indoor depth estimation. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 12663–12673, 2021. 13

    2021

  35. [35]

    DepthFormer: Exploiting long-range correlation and local information for accurate monocular depth estimation.Ma- chine Intelligence Research, 20:837–854, 2023

    Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. DepthFormer: Exploiting long-range correlation and local information for accurate monocular depth estimation.Ma- chine Intelligence Research, 20:837–854, 2023. 1, 13, 17

    2023

  36. [36]

    Binsformer: Revisiting adaptive bins for monocular depth estimation.IEEE Transactions on Image Processing, 33: 3964–3976, 2024

    Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation.IEEE Transactions on Image Processing, 33: 3964–3976, 2024. 13

    2024

  37. [37]

    Unsuper- vised learning reveals interpretable latent representations for translucency perception.PLOS Computational Biology, 19 (2):e1010878, 2023

    Chenxi Liao, Masataka Sawayama, and Bei Xiao. Unsuper- vised learning reveals interpretable latent representations for translucency perception.PLOS Computational Biology, 19 (2):e1010878, 2023. 1

    2023

  38. [38]

    New approaches to 3D vision.Philosophical Transactions of the Royal Society B, 378:20210443, 2023

    Paul Linton, Michael J Morgan, Jenny C A Read, Dhan- raj Vishwanath, Sarah H Creem-Regehr, and Fulvio Domini. New approaches to 3D vision.Philosophical Transactions of the Royal Society B, 378:20210443, 2023. 3

    2023

  39. [39]

    Improving neural network representations using human similarity judgments.Advances in Neural Information Processing Systems, 36:50978–51007,

    Lukas Muttenthaler, Lorenz Linhardt, Jonas Dippel, Robert A Vandermeulen, Katherine Hermann, Andrew Lampinen, and Simon Kornblith. Improving neural network representations using human similarity judgments.Advances in Neural Information Processing Systems, 36:50978–51007,

  40. [40]

    Maxime Oquab, Timoth ´ee Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Rus- sell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang- Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nico- las Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patri...

    Pith/arXiv arXiv 2023

  41. [41]

    ReFusion: 3D reconstruc- tion in dynamic environments for RGB-D cameras exploiting residuals

    Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyrill Stachniss. ReFusion: 3D reconstruc- tion in dynamic environments for RGB-D cameras exploiting residuals. InIEEE/RSJ International Conference on Intelli- gent Robots and Systems, pages 7855–7862. IEEE, 2019. 4, 12

    2019

  42. [42]

    Gloss perception: Searching for a deep neural network that behaves like humans.Journal of Vision, 21(12):14–14,

    Konrad Eugen Prokott, Hideki Tamura, and Roland W Flem- ing. Gloss perception: Searching for a deep neural network that behaves like humans.Journal of Vision, 21(12):14–14,

  43. [43]

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(3):1623–1637, 2020. 2, 13, 14

    2020

  44. [44]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 12179–12188, 2021. 1, 4, 13, 14, 17

    2021

  45. [45]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695. IEEE, 2022. 12

    2022

  46. [46]

    Indoor segmentation and support inference from RGBD images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. InProceedings of the European Conference on Computer Vision, pages 746–760. Springer, 2012. 2, 4, 12

    2012

  47. [47]

    A benchmark for the evalua- tion of RGB-D SLAM systems

    J ¨urgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evalua- tion of RGB-D SLAM systems. InIEEE/RSJ International Conference on Intelligent Robots and Systems, pages 573–

  48. [48]

    Getting aligned on representational alignment.arXiv preprint arXiv:2310.13018, 2023

    Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C Love, Erin Grant, Iris Groen, Jascha Achterberg, et al. Getting aligned on representational alignment.arXiv preprint arXiv:2310.13018, 2023. 1

    Pith/arXiv arXiv 2023

  49. [49]

    SC-Depthv3: Robust self- 10 supervised monocular depth estimation for dynamic scenes

    Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. SC-Depthv3: Robust self- 10 supervised monocular depth estimation for dynamic scenes. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2023. 1, 4, 12, 13, 17

    2023

  50. [50]

    Sparsity invariant CNNs

    Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant CNNs. InInternational Conference on 3D Vision, pages 11–20. IEEE, 2017. 2, 3, 12

    2017

  51. [51]

    Space perception in pictures

    Andrea J van Doorn, Johan Wagemans, Huib de Ridder, and Jan J Koenderink. Space perception in pictures. InSPIE Human Vision and Electronic Imaging XVI, pages 423–430. SPIE, 2011. 3

    2011

  52. [52]

    Lower region: A new cue for figure-ground assign- ment.Journal of Experimental Psychology: General, 131 (2):194–205, 2002

    Shaun P Vecera, Edward K V ogel, and Geoffrey F Wood- man. Lower region: A new cue for figure-ground assign- ment.Journal of Experimental Psychology: General, 131 (2):194–205, 2002. 14

    2002

  53. [53]

    Measuring 3D point configurations in pictorial space

    Johan Wagemans, Andrea J van Doorn, and Jan J Koen- derink. Measuring 3D point configurations in pictorial space. i-Perception, 2(1):77–111, 2011. 3, 4

    2011

  54. [54]

    The metric of visual space.Perception & Psychophysics, 38:483–495, 1985

    Mark Wagner. The metric of visual space.Perception & Psychophysics, 38:483–495, 1985. 7

    1985

  55. [55]

    Mark Wagner and Anthony J. Gambino. Variations in the anisotropy and affine structure of visual space: A geometry of visibles with a third dimension.Topoi, 35(2):583–598,

  56. [56]

    Are deep neural networks adequate behavioral models of human visual per- ception?Annual Review of Vision Science, 9(1):501–524,

    Felix A Wichmann and Robert Geirhos. Are deep neural networks adequate behavioral models of human visual per- ception?Annual Review of Vision Science, 9(1):501–524,

  57. [57]

    Toward practical monocular indoor depth estimation

    Cho-Ying Wu, Jialiang Wang, Michael Hall, Ulrich Neu- mann, and Shuochen Su. Toward practical monocular indoor depth estimation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 3814–3824, 2022. 12, 13

    2022

  58. [58]

    Revealing the dark secrets of masked im- age modeling

    Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked im- age modeling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14475– 14485, 2023. 13

    2023

  59. [59]

    Performance-optimized hierarchical models predict neural responses in higher visual cortex.Proceedings of the Na- tional Academy of Sciences, 111(23):8619–8624, 2014

    Daniel LK Yamins, Ha Hong, Charles F Cadieu, Ethan A Solomon, Darren Seibert, and James J DiCarlo. Performance-optimized hierarchical models predict neural responses in higher visual cortex.Proceedings of the Na- tional Academy of Sciences, 111(23):8619–8624, 2014. 1

    2014

  60. [60]

    Transformer-based attention networks for continuous pixel-wise prediction

    Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. InProceedings of the IEEE/CVF International Conference on Computer vision, pages 16269–16279, 2021. 13, 17

    2021

  61. [61]

    Depth Anything: Unleash- ing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth Anything: Unleash- ing the power of large-scale unlabeled data. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 2, 12, 13, 14

    2024

  62. [62]

    Psychophysical measurement of perceived motion flow of naturalistic scenes.iScience, 26(12):108307, 2023

    Yung-Hao Yang, Taiki Fukiage, Zitang Sun, and Shin’ya Nishida. Psychophysical measurement of perceived motion flow of naturalistic scenes.iScience, 26(12):108307, 2023. 1

    2023

  63. [63]

    HuPerFlow: A comprehensive benchmark for hu- man vs

    Yung-Hao Yang, Zitang Sun, Taiki Fukiage, and Shin’ya Nishida. HuPerFlow: A comprehensive benchmark for hu- man vs. machine motion estimation comparison. InProceed- ings of the Computer Vision and Pattern Recognition Confer- ence, pages 22799–22808, 2025. 1

    2025

  64. [64]

    A statistical explanation of visual space.Nature Neuroscience, 6:632–640, 2003

    Zhiyong Yang and Dale Purves. A statistical explanation of visual space.Nature Neuroscience, 6:632–640, 2003. 3

    2003

  65. [65]

    Metric3d: Towards zero-shot metric 3d prediction from a single image

    Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043–9053, 2023. 13, 14

    2023

  66. [66]

    P 2net: Patch-match and plane-regularization for unsupervised indoor depth esti- mation

    Zehao Yu, Lei Jin, and Shenghua Gao. P 2net: Patch-match and plane-regularization for unsupervised indoor depth esti- mation. InEuropean Conference on Computer Vision, pages 206–222. Springer, 2020. 13

    2020

  67. [67]

    NeW CRFs: Neural window fully-connected CRFs for monocular depth estimation

    Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. NeW CRFs: Neural window fully-connected CRFs for monocular depth estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3906–3915, 2022. 1, 13, 17

    2022

  68. [68]

    GasMono: Geometry-aided self-supervised monocular depth estimation for indoor scenes

    Chaoqiang Zhao, Matteo Poggi, Fabio Tosi, Lei Zhou, Qiyu Sun, Yang Tang, and Stefano Mattoccia. GasMono: Geometry-aided self-supervised monocular depth estimation for indoor scenes. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 16209–16220,

  69. [69]

    To- wards better generalization: Joint depth-pose learning with- out posenet

    Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. To- wards better generalization: Joint depth-pose learning with- out posenet. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151– 9161, 2020. 13

    2020

  70. [70]

    lower-region bias

    Daniel Zoran, Phillip Isola, Dilip Krishnan, and William T. Freeman. Learning ordinal relationships for mid-level vi- sion.Proceedings of the IEEE International Conference on Computer Vision, 2015. 3 11 A. Supplemental Methods All human data and analysis code will be officially released in a public repository upon publication. They are provided in the scr...

    2015