Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Evaluating Robustness of Monocular Depth Estimation with Procedural Scene Perturbations

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read State-of-the-art monocular depth estimators are surprisingly fragile to camera roll and focal-length changes, yet largely robust to lighting changes, according to a new procedurally generated benchmark.

desk verdict A useful new benchmark for monocular-depth robustness, but the headline cross-perturbation difficulty ranking is not yet interpretable because perturbation magnitudes are not matched across types. read the letter →

arxiv 2507.00981 v2 pith:XRONAQMJ submitted 2025-07-01 cs.CV

Jack Nugent , Siyang Wu , Zeyu Ma , Beining Han , Meenal Parakh , Abhishek Joshi , Lingjie Mei , Alexander Raistrick
show 2 more authors
Xinyuan Li Jia Deng
This is my paper
classification cs.CV
keywords monoculardepthestimationrobustnessevaluationproceduralgenerationsyntheticbenchmarkscameraperturbationsself-consistencyaccuracystabilitymodels
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

Monocular depth models are usually judged by accuracy on static benchmarks, which says nothing about whether predictions survive changes to the scene. This paper introduces PDE, a procedurally generated benchmark of 38 indoor and outdoor scenes with 12 perturbation types—camera, object, material, lighting, occlusion, and deformation—totaling 13,693 unique scene variations, and tests nine state-of-the-art depth models on them. Its central claim is that camera roll and dolly zoom are among the most difficult perturbations for the models, much harder than changes to 3D object pose, while lighting changes are among the easiest. The paper further argues that accuracy and robustness are distinct: the most accurate model, DepthPro, is not the most stable, and no model leads on both. If these findings hold, standard benchmark rankings understate how badly depth predictions can warp under camera motion, a central concern for robotics and view synthesis.

What carries the argument

The mechanism is the PDE evaluation pipeline built on Infinigen's procedural generator. A custom layer applies 12 perturbation families that vary exactly one aspect of a scene while holding the rest constant—camera intrinsics (dolly zoom), camera pose (roll, pan/tilt), material (object or scene-wide), lighting, object pose (rotation, translation, resizing), occlusion, non-rigid deformation, and out-of-distribution background swap—yielding 13,693 scene variations across 38 scenes. Robustness is measured in two complementary ways: accuracy stability, the sample variance of the model's error against ground truth over the perturbation set, and self-consistency, the mean squared difference between a perturbed prediction and the base-scene prediction. Self-consistency is only applied when the perturbation leaves ground-truth depth unchanged (or changed only by an SE(2) transform, as for roll), and most metrics are evaluated on the object of interest using eroded masks and standard scale/shift alignment.

What would settle it

Redo the benchmark with perceptually matched perturbation magnitudes—for example, calibrating dolly zoom, roll, and object rotation so human raters judge them equally noticeable—and compare average AbsRel error and stability across perturbation families; the claim that camera perturbations are intrinsically hardest would be falsified if the ordering inverts or disappears under matched magnitudes.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the difficulty of a perturbation does not track its visual severity. Camera roll and dolly zoom alter only the image-formation geometry—the 3D content of the scene is unchanged—yet across the nine models they produce some of the highest error and worst accuracy stability, often by warping an otherwise accurate depth map into physically implausible shapes. Lighting changes, including drastic ones with strong shadows, leave predictions nearly untouched. The paper also establishes a clear accuracy-robustness split: DepthPro obtains the lowest average absolute-relative error on every perturbation type, but UniDepthV2 and Metric3DV2 display the best stability, and UniDepthV2 is the most self-consistent among scale-invariant models. Material swaps and occlusion are additional hard cases, and non-rigid deformations are especially difficult for desks and cabinets. The paper takes these results to show that standard accuracy benchmarks miss a substantial portion of model failure modes.

Load-bearing premise

The ranking of perturbation difficulty assumes the hand-chosen ranges for each perturbation are roughly comparable in stress, since object-pose changes are deliberately kept small while lighting changes are made drastic; if camera-perturbation magnitudes are not matched to object-perturbation magnitudes, the ordering could be an artifact of the chosen ranges rather than a property of the models.

Editorial extensions

If this is right

  • Ranking depth models by standard accuracy alone is misleading, since a model can have the best error on every perturbation type while being less stable than models with worse accuracy.
  • Camera intrinsics and pose changes should be treated as first-class stressors in depth robustness suites, because roll and dolly zoom induce large physically implausible depth warps without changing any scene content.
  • Lighting changes, even drastic ones with strong shadows, are a weak stress test for current depth models.
  • Self-consistency is a stricter criterion than accuracy stability for ground-truth-preserving perturbations: perfect self-consistency implies perfect accuracy stability, but the reverse does not hold.
  • Material swaps and occlusion remain consistently difficult across all tested models, marking them as open problems for robust depth estimation.

Reading between the lines

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

  • Because the paper does not perceptually calibrate perturbation magnitudes across families, a fairer comparison of perturbation difficulty would require matching how noticeable each change is to a human observer; the authors leave this calibration implicit.
  • The paper evaluates perturbations one at a time and notes this in its limitations; combining perturbations, such as camera roll with occlusion or dolly zoom with material swaps, is a natural next experiment that could expose failure modes single perturbations miss.
  • If real-world camera motion produces the same depth warps observed here, training with randomized focal lengths and roll augmentations would be a direct remedy to test; the synthetic-to-real transfer of these robustness patterns remains open.
  • The appendix's finding that eroding object masks narrows the gap between models suggests edge prediction errors drive average error more than robustness, implying that fixing edges may improve accuracy without improving stability.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces PDE (Procedural Depth Evaluation), a benchmark for robustness of monocular depth estimation built on Infinigen scenes. It defines 12 controlled perturbation types spanning camera intrinsics, camera pose, object pose/resizing/occlusion/deformation, materials, lighting, and OOD backgrounds, and evaluates 9 recent models under AbsRel, δ, and RMSE metrics with accuracy-stability and self-consistency measures. The central claim is that camera perturbations, especially dolly zoom and roll, are among the most challenging for all models, while lighting and small object-pose changes are relatively easy, and that no single model dominates in both accuracy and robustness.

Significance. The paper is a potentially valuable evaluation resource: it goes beyond image-level corruptions such as RoboDepth to controlled 3D scene perturbations, uses high-quality procedural ground truth, evaluates a broad set of models, and releases code and data. The accuracy-stability versus self-consistency distinction is a useful framework, and the finding that camera intrinsics and pose changes can induce large metric distortions would, if robust, be an important caveat for standard depth benchmarks. However, the headline comparative ranking across perturbation types is the main contribution, and it currently relies on intentionally asymmetric and unreported perturbation magnitudes, so the significance is conditional on addressing that issue.

major comments (3)
  1. [§3.1, Table 2, Figure 3] The central ranking that camera dolly zoom and roll are 'much more difficult' than object-pose perturbations is not interpretable without matching perturbation stress across types. Section 3.1 explicitly says that perturbations that do not change ground truth, such as lighting and material swaps, 'can be made drastic,' while perturbations that change ground truth, such as object rotation and translation, are 'limited to a small neighborhood.' Dolly zoom changes ground-truth depth, yet no analogous small-neighborhood constraint is stated for it, and camera roll is a ground-truth-preserving rotation that can be made arbitrarily large. The paper does not report the actual parameter ranges (only 'up to 60 different parameter settings') or any calibration in terms of perceptual change, ground-truth depth change, or another task-relevant stress measure. Consequently, the ranking in Table 2 and Figure 3 may reflect the chosen ranges rather than model properties. Please fix this by reporting per-perturbation magnitude sweeps, defining and applying a comparable stress measure, or explicitly rephrasing the claim as holding 'at the chosen perturbation magnitudes.'
  2. [§4.2, Table 2, Figure 3] The cross-model and cross-perturbation claims lack uncertainty quantification. All conclusions are point estimates aggregated over 38 scenes (8 per object category), with no per-scene variance, confidence intervals, or significance tests. Given the scene-to-scene variability visible in the paper's own qualitative examples, the gaps in Table 2—for instance Cam Dolly Zoom versus Object Rotation (Avg. Err. 1.94 vs 1.74; stability 0.55 vs 0.31)—may or may not be statistically robust. Please provide per-scene distributions, error bars, or paired statistical comparisons across scenes before asserting that a perturbation type is 'among the most challenging.'
  3. [§3.2, Appendix A] The self-consistency metric κ is defined as the average squared difference between the perturbed prediction and the base prediction, but the formula writes κ = (1/N) Σ Δ(x_i, x_0)^2, where Δ is already an error metric such as AbsRel. This makes κ a squared version of a metric-dependent distance; please clarify whether the squaring is intended and, if so, state it explicitly. More importantly, the alignment for self-consistency normalizes the base prediction to a median depth of 1 and then aligns to that normalized prediction, which makes κ dependent on the model's own base output and on an arbitrary normalization. The paper should state clearly that self-consistency results in Table 3 are only comparable among the metric-depth models and should not be used to rank affine-invariant models.
minor comments (5)
  1. [Throughout] There are several typos and grammatical slips: 'evalutes' in Section 2, 'pertubations' in the Table 1 caption, 'adversial' in Section 2, and 'Top models shows' in the Figure 3 caption.
  2. [§3.1, Table 1] For Cam Dolly Zoom, marking 'Depth Image Modulo (SE(2))' as Y is confusing: changing focal length while translating the camera changes perspective projection for a 3D object (near and far surfaces scale differently), so the object's depth image is not merely an SE(2) transform of the original. Please clarify or correct the table entry.
  3. [Appendix C] The definition of δ0.125 is not typeset correctly: the expression 'max(ˆyi/yi, yi/ˆyi) < 1.250.125' should presumably be max(ˆyi/yi, yi/ˆyi) < 1.25^0.125; please fix the formatting.
  4. [§3.3, Appendix A] The paper does not specify which models use scale-only alignment and which use scale-and-shift alignment; the main text says 'standard procedures as in Marigold and MiDaS,' but a per-model list would improve reproducibility.
  5. [Figure 2] The caption says models are ranked by several metrics 'on the object of interest, and by AbsRel on the full scene including background'; please clarify which panels correspond to which metric and mask, since the visual grouping is not self-explanatory.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the benchmark measurements are evaluated against external models with standard alignment, and no claimed result reduces to a fitted parameter or self-citation.

full rationale

This paper is an empirical evaluation benchmark, not a derivation of predictions from inputs. The central findings—ranking perturbations by difficulty and ranking models by robustness—are obtained by running nine externally developed depth models on procedurally generated scenes and measuring standard errors (AbsRel, δ, RMSE) after conventional scale/shift alignment. No parameter is fitted to the data and then reported as a prediction; the scale/shift alignment is a standard evaluation procedure applied identically across perturbations and does not force any particular ranking. The use of Infinigen [25, 26], cited from the authors' own prior work, is a tool for scene synthesis and is not load-bearing in a circular sense: the models being evaluated were not trained on Infinigen data, and the robustness measurements are externally reproducible once the code and scenes are released. The reader-flagged concern that perturbation magnitudes are not matched across types (e.g., lighting made 'drastic' while object pose kept in a 'small neighborhood') is a legitimate threat to the interpretation of the comparative difficulty ranking, but it is a benchmark-design validity issue, not circularity: the observed errors are measured, not constructed from the choice of ranges. The paper's own limitations section (Section 5) acknowledges coverage gaps, but does not hide any circular step. Accordingly, no circular step can be quoted or exhibited, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims of the paper are empirical findings from a benchmark. The main free parameters are the perturbation ranges chosen by the authors, which directly influence the relative difficulty ranking. The axioms are the standard assumptions of using synthetic data and alignment procedures for depth evaluation.

free parameters (3)
  • Perturbation magnitude ranges per type = Not specified numerically
    Object rotation and translation are limited to a small neighborhood, occlusion is 10-40%, and lighting is described as drastic (Section 3.1). These author-chosen ranges affect the relative difficulty ranking across perturbation types, which is the central finding.
  • Number of scenes per object category = 8 scenes per category, 38 total
    Coverage choice; findings could shift with more scenes or categories (Section 3).
  • Object mask erosion = 1 pixel for main results; up to larger for analysis
    Erosion affects alignment and error metrics, particularly for thin objects; Appendix B.2 shows results change with erosion level.
assumptions (4)
  • domain assumption Infinigen renders high-quality ground truth depth from synthetic scenes.
    Used throughout; accuracy of ground truth is assumed sufficient for benchmarking (Section 3).
  • domain assumption The evaluated models were not trained on Infinigen data.
    Stated in Section 1; if a model had seen Infinigen-like data during training, the unbiased evaluation claim would weaken.
  • domain assumption Human perception can reliably judge depth in Infinigen scenes, making the domain gap acceptable.
    Section 1 asserts 'humans can still easily perceive their depth', but this is not empirically verified.
  • domain assumption Evaluating after scale/shift alignment is the correct way to compare models with different output scales.
    Standard in depth evaluation, but alignment can absorb some robustness failures; the paper reports both scale and scale+shift variants.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Robustness of Monocular Depth Estimation with Procedural Scene Perturbations." pith.science (2026). https://pith.science/paper/XRONAQMJ

@misc{pith2026250700981,
  author       = {Pith},
  title        = {Pith review of: Evaluating Robustness of Monocular Depth Estimation with Procedural Scene Perturbations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XRONAQMJ}},
  note         = {Machine review of arXiv:2507.00981}
}
read the original abstract

Recent years have witnessed substantial progress on monocular depth estimation, particularly as measured by the success of large models on standard benchmarks. However, performance on standard benchmarks does not offer a complete assessment, because most evaluate accuracy but not robustness. In this work, we introduce PDE (Procedural Depth Evaluation), a new benchmark which enables systematic robustness evaluation. PDE uses procedural generation to create 3D scenes that test robustness to various controlled perturbations, including object, camera, material and lighting changes. Our analysis yields interesting findings on what perturbations are challenging for state-of-the-art depth models, which we hope will inform further research. Code and data are available at https://github.com/princeton-vl/proc-depth-eval.

Figures

Figures reproduced from arXiv: 2507.00981 by the authors.

Figure 1
Figure 1. Perturbation Types We use procedural generation to construct controlled perturbations which vary one aspect of a scene while keeping others constant. We show a single example for each perturbation type, with the object of interest highlighted by red boxes. for further research. Our work aims to provide similar insights for the development of robust depth estimation models. Robustness evaluation with procedural pertu… view at source ↗
Figure 2
Figure 2. Depth estimation models ranked by Average Error, Accuracy (In)Stability and Self- [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Variations ranked by Average Error, Accuracy (In)Stability, Self-(In)Consistency on the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Camera Pose Robustness model exhibits strong robustness to these variations, perhaps indicating a success of their large scale training dataset of “pseudo-labeled” real images [42] [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Non-Rigid deformation perfor￾mance by object category. Non-rigid Object Deformation In contrast to material alterations, the non-rigid object deformation variation cre￾ates scenes that differ in semantically meaningful but com￾plex manners. Depth Pro performs the best …
Figure 7
Figure 7. Figure 7: OOD Deformation Robustness – We visualize point clouds from DepthPro and UniDepthV2 for a desk with an unusual added bump. DepthPro attempts to reconstruct it whereas UniDepth predicts a mostly flat surface. physically implausible since the chair seat is connected to t…
Figure 8
Figure 8. Figure 8: Model performance by FLOPs. Lines indicate a “family” of models (e.g. UniDepthV2- [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 10
Figure 10. Figure 10: Impact of object mask erosion on model performance. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Geometric Distillation from Rectified Stereo: Leveraging Epipolar Cues for Monocular Depth

    cs.CV 2026-07 conditional novelty 7.0 of 10

    EpiDistill uses depth-guided epipolar attention and learnable rectified stereo tokens to distill multi-view scale knowledge into single-view monocular depth models.

Reference graph

Works this paper leans on

47 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Muller

    S. Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Muller. Zoedepth: Zero-shot transfer by com- bining relative and metric depth. ArXiv, abs/2302.12288, 2023. URL https://api.semanticscholar. org/CorpusID:257205739

    arXiv 2023

  2. [2]

    Midas v3.1 - a model zoo for robust monocular rel- ative depth estimation

    Reiner Birkl, Diana Wofk, and Matthias Müller. Midas v3.1 - a model zoo for robust monocular rel- ative depth estimation. ArXiv, abs/2307.14460, 2023. URL https://api.semanticscholar.org/ CorpusID:260202837

    arXiv 2023

  3. [3]

    Richter, and Vladlen Koltun

    Alexey Bochkovskiy, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second.ArXiv, abs/2410.02073,

    arXiv

  4. [4]

    D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV , LNCS 7577, pages 611–625. Springer-Verlag, October 2012

    work page 2012

  5. [5]

    Procthor: Large-scale embodied ai using procedural generation

    Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, and Roozbeh Mottaghi. Procthor: Large-scale embodied ai using procedural generation. ArXiv, abs/2206.06994, 2022. URL https://api.semanticscholar. org/CorpusID:249642405

    arXiv 2022

  6. [6]

    Meta-sim2: Unsupervised learning of scene structure for synthetic data generation

    Jeevan Devaranjan, Amlan Kar, and Sanja Fidler. Meta-sim2: Unsupervised learning of scene structure for synthetic data generation. In European Conference on Computer Vision, 2020. URL https://api. semanticscholar.org/CorpusID:221186741

    work page 2020

  7. [7]

    3d-front: 3d furnished rooms with layouts and semantics

    Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Cao Li, Zengqi Xun, Chengyue Sun, Yiyun Fei, Yu qiong Zheng, Ying Li, Yi Liu, Peng Liu, Lin Ma, Le Weng, Xiaohang Hu, Xin Ma, Qian Qian, Rongfei Jia, Binqiang Zhao, and Hao Helen Zhang. 3d-front: 3d furnished rooms with layouts and semantics. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pa...

    work page 2021

  8. [8]

    Adversarial robustness for visual grounding of multimodal large language models

    Kuofeng Gao, Yang Bai, Jiawang Bai, Yong Yang, and Shu-Tao Xia. Adversarial robustness for visual grounding of multimodal large language models. ArXiv, abs/2405.09981, 2024. URL https://api. semanticscholar.org/CorpusID:269791017

    arXiv 2024

Show all 47 references
  1. [9]

    Vision meets robotics: The kitti dataset

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research , 32:1231 – 1237, 2013. URL https://api. semanticscholar.org/CorpusID:9455111

    2013

  2. [10]

    Goodfellow, Jonathon Shlens, and Christian Szegedy

    Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2014. URL https://api.semanticscholar.org/CorpusID:6706414

    2014 arXiv

  3. [11]

    Evaluating concurrent robust- ness of language models across diverse challenge sets

    Vatsal Gupta, Pranshu Pandya, Tushar Kataria, Vivek Gupta, and Dan Roth. Evaluating concurrent robust- ness of language models across diverse challenge sets. In Conference on Empirical Methods in Natural Language Processing, 2023. URL https://api.semanticscholar.org/CorpusID:265212792

    2023

  4. [12]

    Natural adversarial examples

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Xiaodong Song. Natural adversarial examples. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15257–15266, 2019. URL https://api.semanticscholar.org/CorpusID:196831327

    2021

  5. [13]

    Xiaoyan Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen

    Mu Hu, Wei Yin, China. Xiaoyan Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Patte...

    2024

  6. [14]

    Champion-level drone racing using deep reinforcement learning

    Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Champion-level drone racing using deep reinforcement learning. Nature, 620:982 – 987,

  7. [15]

    Repurposing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 9492–9502, 2023. U...

    2024

  8. [16]

    Czarnecki

    Samin Rahman Khan, Buu Phan, Rick Salay, and K. Czarnecki. Procsy: Procedural synthetic dataset generation towards influence factor studies of semantic segmentation networks. In CVPR Workshops, 2019. URL https://api.semanticscholar.org/CorpusID:198166884

    2019

  9. [17]

    Cottereau, and Wei Tsang Ooi

    Lingdong Kong, Shaoyuan Xie, Hanjiang Hu, Lai Xing Ng, Benoit R. Cottereau, and Wei Tsang Ooi. Robodepth: Robust out-of-distribution depth estimation under corruptions, 2023. URL https://arxiv. org/abs/2310.15171

    2023 arXiv

  10. [18]

    Goodfellow, and Samy Bengio

    Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. ArXiv, abs/1607.02533, 2016. URL https://api.semanticscholar.org/CorpusID:1257772

    2016 arXiv

  11. [19]

    Pulling things out of perspective

    Lubor Ladicky, Jianbo Shi, and Marc Pollefeys. Pulling things out of perspective. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89–96, 2014

    2014

  12. [20]

    Learning quadrupedal locomotion over challenging terrain

    Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning quadrupedal locomotion over challenging terrain. Science Robotics , 5, 2020. URL https://api. semanticscholar.org/CorpusID:224828219

    2020

  13. [21]

    Spring: A high- resolution high-detail dataset and benchmark for scene flow, optical flow and stereo

    Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andrés Bruhn. Spring: A high- resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4981–4991, ...

    2023

  14. [22]

    Unidepth: Universal monocular metric depth estimation

    Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10106–10116, 2024. URL https://api...

    2024

  15. [23]

    UniDepthV2: Universal monocular metric depth estimation made simpler, 2025

    Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. UniDepthV2: Universal monocular metric depth estimation made simpler, 2025. URL https: //arxiv.org/abs/2502.20110

    2025 arXiv

  16. [24]

    Visual adversarial examples jailbreak aligned large language models

    Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In AAAI Conference on Artificial Intelligence, 2023. URL https://api.semanticscholar.org/CorpusID:259244034

    2023

  17. [25]

    Alexander R. E. Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, Alejandro Newell, Hei Law, Ankit Goyal, Kaiyu Yang, and Jia Deng. Infinite photorealistic worlds using procedural generation. 2023 IEEE/C...

    2023

  18. [26]

    Alexander R. E. Raistrick, Lingjie Mei, Karhan Kayan, David Yan, Yiming Zuo, Beining Han, Hongyu Wen, Meenal Parakh, Stamatis Alexandropoulos, Lahav Lipson, Zeyu Ma, and Jia Deng. Infinigen indoors: Photorealistic indoor scenes using procedural generation. 2024 IEEE/CVF Confer...

    2024

  19. [27]

    Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer, 2020

    René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer, 2020. URL https: //arxiv.org/abs/1907.01341

    2020 arXiv

  20. [28]

    Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding

    Mike Roberts and Nathan Paczan. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10892–10902,

    2021

  21. [29]

    Make3d: Learning 3d scene structure from a single still image

    Ashutosh Saxena, Min Sun, and Andrew Y Ng. Make3d: Learning 3d scene structure from a single still image. IEEE transactions on pattern analysis and machine intelligence, 31(5):824–840, 2008

    2008

  22. [30]

    Schönberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger

    Thomas Schöps, Johannes L. Schönberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi- camera videos. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  23. [31]

    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. In European Conference on Computer Vision, 2012. URL https://api. semanticscholar.org/CorpusID:545361. 12

    2012

  24. [32]

    Erhan, Ian J

    Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, D. Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. CoRR, abs/1312.6199, 2013. URL https: //api.semanticscholar.org/CorpusID:604334

    2013 arXiv

  25. [33]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras

    Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In Neural Information Processing Systems, 2021. URL https://api.semanticscholar.org/CorpusID: 237278112

    2021

  26. [34]

    Deep patch visual odometry

    Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. ArXiv, abs/2208.04726, 2022. URL https://api.semanticscholar.org/CorpusID:251442450

    2022 arXiv

  27. [35]

    Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z

    Igor Vasiljevic, Nicholas I. Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R. Walter, and Gregory Shakhnarovich. Diode: A dense indoor and outdoor depth dataset. ArXiv, abs/1908.00463, 2019. UR...

    1908 arXiv

  28. [36]

    Jia, Bo Li, and Jingjing Liu

    Boxin Wang, Shuohang Wang, Yu Cheng, Zhe Gan, R. Jia, Bo Li, and Jingjing Liu. Infobert: Improving robustness of language models from an information theoretic perspective. ArXiv, abs/2010.02329, 2020. URL https://api.semanticscholar.org/CorpusID:222140859

    2010 arXiv

  29. [37]

    Boxin Wang, Chejian Xu, Shuohang Wang, Zhe Gan, Yu Cheng, Jianfeng Gao, Ahmed Hassan Awadallah, and B. Li. Adversarial glue: A multi-task benchmark for robustness evaluation of language models. ArXiv, abs/2111.02840, 2021. URL https://api.semanticscholar.org/CorpusID:242757097

    2021 arXiv

  30. [38]

    Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision

    Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. ArXiv, abs/2410.19115, 2024. URL https://api.semanticscholar.org/Corpus...

    arXiv

  31. [39]

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian A. Scherer. Tartanair: A dataset to push the limits of visual slam. 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 4909–491...

    2020

  32. [40]

    Synscapes: A photorealistic synthetic dataset for street scene parsing

    Magnus Wrenninge and Jonas Unger. Synscapes: A photorealistic synthetic dataset for street scene parsing. ArXiv, abs/1810.08705, 2018. URL https://api.semanticscholar.org/CorpusID:53047282

    2018 arXiv

  33. [41]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371–10381, 2024. URL https://api.semanti...

    2024

  34. [42]

    Depth anything v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. ArXiv, abs/2406.09414, 2024. URL https://api.semanticscholar. org/CorpusID:270440448

    2024 arXiv

  35. [43]

    Holodeck: Language guided generation of 3d embodied ai environments

    Yue Yang, Fan-Yun Sun, Luca Weihs, Eli VanderBilt, Alvaro Herrasti, Winson Han, Jiajun Wu, Nick Haber, Ranjay Krishna, Lingjie Liu, Chris Callison-Burch, Mark Yatskar, Aniruddha Kembhavi, and Christopher Clark. Holodeck: Language guided generation of 3d embodied ai environment...

    2024

  36. [44]

    object” setting or on all pixels not in the background for the “full scene

    Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. ArXiv, abs/2305.16934, 2023. URL https://api.semanticscholar.org/CorpusID:258947177. 13 A Evaluation Details For eac...

    2023 arXiv

  37. [2020]

    URL https://api.semanticscholar.org/CorpusID:226254406

  38. [2023]

    URL https://api.semanticscholar.org/CorpusID:261357832

  39. [2024]

    URL https://api.semanticscholar.org/CorpusID:273098591

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.