Pith. sign in

REVIEW 4 major objections 6 minor 55 references

CoL3D: Collaborative Learning of Single-view Depth and Camera Intrinsics for Metric 3D Shape Recovery

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CoL3D claims that depth and camera intrinsics can be learned together in one network, and that doing so improves metric 3D shape recovery from a single image beyond methods that assume known camera settings.

desk verdict A sensible joint depth-intrinsics framework whose headline depth gains are not controlled against its own baseline, so the central claim is weaker than the paper suggests. read the letter →

arxiv 2502.08902 v1 pith:CCGWWPBB submitted 2025-02-13 cs.CV

Chenghao Zhang , Lubin Fan , Shen Cao , Bojian Wu , Jieping Ye This is my paper
classification cs.CV MSC 68T45
keywords monoculardepthestimationcameraintrinsicscalibrationincidencefieldmetric3DshaperecoverypointcloudChamferdistancesingle-imagereconstructioncollaborativelearning
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

CoL3D claims that monocular depth and camera intrinsics should be learned together in a single network, not in separate pipelines. The paper argues that depth is a 3D prior for estimating intrinsics: given depth and the known Euclidean sizes of reference objects, the 4-DoF camera intrinsics are determined. It then trains one shared encoder-decoder to predict both a depth map and an incidence field, and optimizes depth, intrinsics, and the resulting point cloud jointly. If the claim holds, a robot can recover metric 3D shape from one RGB image without knowing the focal length or principal point, and does so more accurately than pipelines that rely on ground-truth intrinsics at test time.

What carries the argument

The load-bearing mechanism is the incidence field, a per-pixel field of rays from 3D points to 2D pixels that parameterizes the camera intrinsics. CoL3D fixes a canonical incidence field built from a canonical focal length and image-center principal point, then makes the camera head predict only the residual field between the canonical and ground-truth fields, keeping the unprojection from intrinsics to 3D points fully differentiable. The second mechanism is the shape similarity loss, the Chamfer distance between predicted and ground-truth point clouds, which couples depth and intrinsics errors in 3D space so that per-pixel depth errors that would distort scene shape are penalized. A shared encoder-decoder provides the collaborative channel through which depth gradients improve calibration and calibration gradients improve depth.

What would settle it

Train CoL3D on a dataset where RGB and depth are captured by a wide-angle camera with known radial distortion, while labels are generated assuming a pinhole model; if the mean focal-length (FoV) error degrades sharply relative to distortion-free benchmarks while a pinhole-only baseline does not, the ideal-pinhole assumption is load-bearing. Alternatively, synthetically apply barrel distortion to NYU images and keep pinhole depth labels: the proof's constraint no longer holds for observed pixels, so the predicted FoV error should increase if the proof is what makes the method work.

Watch

Extended reading notes

Core claim

The central discovery is that depth estimation and camera calibration reinforce each other when trained collaboratively at three levels: depth, camera, and point cloud. The paper proves, under an ideal pinhole model, that four non-overlapping pixel groups with known 3D Euclidean distances are enough to determine the 4-DoF intrinsics from a depth map, making depth a constraint on calibration rather than merely a consumer of it. CoL3D embodies this by predicting a residual incidence field around a canonical camera and by adding a Chamfer-distance loss between predicted and ground-truth point clouds. The experimental claim is that this cooperative setup outperforms state-of-the-art in-domain monocular depth methods on NYU and KITTI while also estimating accurate intrinsics, and that it yields better metric 3D shapes than methods that use ground-truth intrinsics when unprojecting their depth predictions.

Load-bearing premise

The whole argument relies on an ideal pinhole camera with no lens distortion and on ground-truth depth and intrinsics being perfectly aligned with the RGB image; if real images violate either condition, the unprojection geometry used in both the proof and the Chamfer loss is biased, and the collaborative training signal is corrupted.

Editorial extensions

If this is right

  • Monocular depth networks can output camera intrinsics as a by-product, eliminating the need for known focal length at deployment.
  • Adding a point-cloud-level loss makes depth predictions more shape-faithful than per-pixel depth losses alone, with fewer 3D outliers.
  • The canonical-field residual design keeps calibration stable across datasets with different lens properties.
  • In-domain training on a single dataset can reach or beat state-of-the-art depth accuracy, so customized robot-perception settings can avoid massive multi-dataset pretraining.

Reading between the lines

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

  • If depth can constrain calibration, then pseudo-depth from any monocular model could serve as weak supervision for calibrating cameras in datasets that lack intrinsics; the paper demonstrates this only in its street-view experiment and does not claim it as a general recipe.
  • The canonical residual idea may extend to lens distortion, with a network predicting residual distortion coefficients around a canonical distortion-free model, although the paper's pinhole proof does not cover that case.
  • Because the Chamfer loss treats the whole scene shape as the optimization target, the framework may be sensitive to dynamic objects or misalignment between depth and RGB capture; an ablation with misaligned labels would test this limit.
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

4 major / 6 minor

Summary. The paper proposes CoL3D, a unified network that jointly estimates monocular depth and camera intrinsics from a single RGB image, motivated by a theoretical proposition that depth constrains the 4-DoF intrinsics given known 3D distances between selected pixel groups. The method introduces a canonical incidence field as a residual-learning prior and a Chamfer-distance-based shape loss in the point cloud space. Experiments on NYU-Depth-v2, KITTI, SUN RGB-D, GSV, and Taskonomy report strong in-domain depth estimation, camera calibration performance comparable to single-task methods, and improved 3D point-cloud quality relative to depth-only baselines that use ground-truth intrinsics.

Significance. If the empirical claims hold, the paper demonstrates a practical way to recover metric 3D shapes without requiring camera intrinsics at inference, which is valuable for robotics and embodied AI. The canonical incidence field is a sensible inductive bias for camera calibration, and the point-cloud loss is a natural objective for 3D shape quality. The theoretical proposition, while not used as a hard constraint in the optimization, is a useful conceptual contribution. However, the paper's central empirical claim that collaborative learning is responsible for the gains is not fully established because the key ablation shows only marginal depth improvement and the primary 3D metric coincides with the training loss.

major comments (4)
  1. [Eq. (6), Section IV-C] The cosine similarity loss as printed is dimensionally invalid: (V_i · V_cano)^T V*_i is a scalar multiplied by a vector, so the right-hand side is a sum of vectors, not a scalar loss. This makes the actual optimization objective undefined. Please replace Eq. (6) with a proper scalar cosine-similarity loss, e.g., 1 - (V_i · V*_i)/(|V_i||V*_i|), or state the intended elementwise operation explicitly.
  2. [Section V-E, Table VII vs. Table I] The ablation in Table VII shows that the full model achieves RMSE 0.294 on NYU while the 'w/o Camera Head' depth-only baseline achieves 0.295, a difference of 0.3%. Yet Table I reports a 6% RMSE improvement over iDisc and IEBins. Without a controlled single-task baseline using the identical training protocol (backbone, resolution, iterations, augmentation, and loss schedule), the headline depth gains cannot be attributed to the proposed collaborative modules. The paper should present a direct depth-only baseline under the same protocol as the main comparison.
  3. [Section V-D, Table VI and Eq. (8)] The primary 3D shape metric, Chamfer distance DCham, is exactly the loss Lcd that CoL3D minimizes, while all compared baselines are trained with per-pixel depth losses and use ground-truth intrinsics for unprojection. This creates an expected advantage for CoL3D that is independent of any depth-intrinsics collaboration. To support the claim that collaborative learning improves 3D shape quality, the paper should compare against a baseline trained with the same Chamfer loss but without the camera head (or with ground-truth intrinsics), and should report additional 3D metrics that are not part of the training objective.
  4. [Section V-E (no KITTI ablation)] The paper claims state-of-the-art depth and 3D shape results on KITTI (Table III and Fig. 4), but the component ablation is reported only for NYU (Table VII) and Taskonomy (Table IV). Without a KITTI ablation isolating the contributions of Vcano and Lcd, the effectiveness of the collaborative mechanism on outdoor scenes remains unsupported. Please add an ablation study on KITTI covering depth, FoV error, and 3D shape metrics.
minor comments (6)
  1. [Appendix VII, Proposition proof] The proof states that stacking four quadratic equations of the form of Eq. (17) yields a solution via Levenberg-Marquardt, but it does not address whether the solution is unique or what conditions on the pixel groups guarantee a unique determination. A brief argument about genericity or a reference to a minimal-solver analysis would strengthen the theoretical claim.
  2. [Table I] The UniDepth row reports A.Rel = 0.626 and RMSE = 0.232; these numbers appear inconsistent with each other and with published UniDepth results. Please verify the values and add a footnote explaining the evaluation protocol for that row.
  3. [Table V] The labels 'Ours w/o Asm.' and 'Ours w Asm.' are not defined in the text or caption. Please define 'Asm.' (presumably an assumption or auxiliary component) and specify what is included in each variant.
  4. [Section IV-A] The equation 'V_res · V_cano = V_gt' uses the symbol '·' without defining whether it denotes elementwise multiplication, matrix multiplication, or some other operation. Please clarify the intended composition rule for the residual incidence field.
  5. [Figure 1] The figure lists 'No collected depth map' as an attribute of the proposed pipeline, but the method uses ground-truth depth for training on NYU, KITTI, and Taskonomy. The caption should clarify that this refers to inference-time requirements only, not to the training protocol.
  6. [Section IV-B, Eq. (4)] The Chamfer distance is defined over dense depth maps containing hundreds of thousands of points, and a naive implementation would be computationally prohibitive. Please specify whether the loss is computed on a subsampled set of points and, if so, describe the sampling procedure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Proposition is a standalone proof, the losses are direct supervision against ground truth, and there are no load-bearing self-citations; the Lcd/DCham identity is an experimental-control caveat, not a circular step.

full rationale

The claimed first-principles result (the Proposition in Section III and its proof in Appendix VII) is a standalone mathematical statement: given a depth map and four 3D Euclidean distances of pixel groups, the 4-DoF intrinsic matrix can be solved. The proof starts from the pinhole unprojection equation and does not assume the conclusion; it is used as motivation for the collaborative framework, not as a fitted constraint. The method is straightforwardly supervised: depth is trained with Lsilog against ground-truth depth, the incident field with Lcos against ground-truth incidence, and the point cloud with Lcd against ground-truth point clouds. The canonical incidence field is a fixed reparameterization (Vcano) providing a residual-learning prior; it does not smuggle in the target result. There are no load-bearing self-citations: references to iDisc, Metric3D, UniDepth, and the incidence field are external prior works. The GSV experiment uses depth pseudo-labels from an external off-the-shelf model (Metric3D) as supervision; this is ordinary transfer, not circular, because the calibration result is not derived from CoL3D's own outputs. One methodological caveat, not a circularity: the Chamfer loss Lcd in Eq. (8) equals the DCham metric reported in Table VI, so the external 3D-shape comparison is not controlled for the training objective; this is an experimental-design concern and does not make the derivation self-referential.

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

The method does not introduce new physical entities. It relies on standard pinhole camera assumptions, supervised training data, and a small number of hand-tuned hyperparameters.

free parameters (3)
  • Loss weights alpha, beta, gamma = 1, 10, 1
    Hand-chosen in Eq. (9) to balance depth, incidence, and Chamfer losses; not tuned per dataset.
  • Canonical focal length fc = Not disclosed
    Used to construct Vcano in Eq. (3); the authors state results are insensitive to fc but do not report its value or range.
  • Scale-invariant loss parameter lambda = 0.5
    Set in Eq. (5) following prior work; a mild hyperparameter.
assumptions (3)
  • domain assumption Pinhole camera model with zero distortion
    Used in Eq. (1) and the Proposition; real cameras often have radial distortion, which is not modeled.
  • domain assumption Ground truth depth and intrinsics are available for supervision
    The method trains on datasets with depth and camera intrinsics labels; without them the collaborative losses cannot be formed.
  • domain assumption Known Euclidean distances between pixel groups in the Proposition
    The theoretical result assumes 4 known 3D distances; these require external metric information about the scene.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoL3D: Collaborative Learning of Single-view Depth and Camera Intrinsics for Metric 3D Shape Recovery." pith.science (2026). https://pith.science/paper/CCGWWPBB

@misc{pith2026250208902,
  author       = {Pith},
  title        = {Pith review of: CoL3D: Collaborative Learning of Single-view Depth and Camera Intrinsics for Metric 3D Shape Recovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CCGWWPBB}},
  note         = {Machine review of arXiv:2502.08902}
}
read the original abstract

Recovering the metric 3D shape from a single image is particularly relevant for robotics and embodied intelligence applications, where accurate spatial understanding is crucial for navigation and interaction with environments. Usually, the mainstream approaches achieve it through monocular depth estimation. However, without camera intrinsics, the 3D metric shape can not be recovered from depth alone. In this study, we theoretically demonstrate that depth serves as a 3D prior constraint for estimating camera intrinsics and uncover the reciprocal relations between these two elements. Motivated by this, we propose a collaborative learning framework for jointly estimating depth and camera intrinsics, named CoL3D, to learn metric 3D shapes from single images. Specifically, CoL3D adopts a unified network and performs collaborative optimization at three levels: depth, camera intrinsics, and 3D point clouds. For camera intrinsics, we design a canonical incidence field mechanism as a prior that enables the model to learn the residual incident field for enhanced calibration. Additionally, we incorporate a shape similarity measurement loss in the point cloud space, which improves the quality of 3D shapes essential for robotic applications. As a result, when training and testing on a single dataset with in-domain settings, CoL3D delivers outstanding performance in both depth estimation and camera calibration across several indoor and outdoor benchmark datasets, which leads to remarkable 3D shape quality for the perception capabilities of robots.

Figures

Figures reproduced from arXiv: 2502.08902 by the authors.

Figure 1
Figure 1. Comparison of our collaborative learning framework with single [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed CoL3D framework. It consists of an Encoder and Decoder for latent feature extraction, a Depth Head for depth [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative 3d shape comparison on the NYU dataset. The red boxes indicate the regions to focus on. Input iDisc Ours [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative 3D shape comparison on the KITTI dataset. The red boxes show the regions to focus on. D. 3D Shape Recovery Tab. VI shows the performance comparison results of 3D shape recovery quality on NYU with other single-task MDE methods. We report 3D metrics includin…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 49 canonical work pages

  1. [1]

    3dp3: 3d scene perception via probabilistic programming,

    N. Gothoskar, M. Cusumano-Towner, B. Zinberg, M. Ghavamizadeh, F. Pollok, A. Garrett, J. Tenenbaum, D. Gutfreund, and V . Mans- inghka, “3dp3: 3d scene perception via probabilistic programming,” in NeurIPS, 2021, pp. 9600–9612

    work page 2021

  2. [2]

    Self- supervised pre-training with masked shape prediction for 3d scene understanding,

    L. Jiang, Z. Yang, S. Shi, V . Golyanik, D. Dai, and B. Schiele, “Self- supervised pre-training with masked shape prediction for 3d scene understanding,” in CVPR, 2023, pp. 1168–1178

    work page 2023

  3. [3]

    Adabins: Depth estimation using adaptive bins,

    S. F. Bhat, I. Alhashim, and P. Wonka, “Adabins: Depth estimation using adaptive bins,” in CVPR, 2021, pp. 4009–4018

    work page 2021

  4. [4]

    Neural window fully- connected crfs for monocular depth estimation,

    W. Yuan, X. Gu, Z. Dai, S. Zhu, and P. Tan, “Neural window fully- connected crfs for monocular depth estimation,” in CVPR, 2022, pp. 3916–3925

    work page 2022

  5. [5]

    Depthformer: Exploiting long- range correlation and local information for accurate monocular depth estimation,

    Z. Li, Z. Chen, X. Liu, and J. Jiang, “Depthformer: Exploiting long- range correlation and local information for accurate monocular depth estimation,” Machine Intelligence Research , vol. 20, no. 6, pp. 837– 854, 2023

    work page 2023

  6. [6]

    idisc: Internal discretization for monocular depth estimation,

    L. Piccinelli, C. Sakaridis, and F. Yu, “idisc: Internal discretization for monocular depth estimation,” in CVPR, 2023, pp. 21 477–21 487

    work page 2023

  7. [7]

    Iebins: Iterative elastic bins for monocular depth estimation,

    S. Shao, Z. Pei, X. Wu, Z. Liu, W. Chen, and Z. Li, “Iebins: Iterative elastic bins for monocular depth estimation,” in NeurIPS, 2023

    work page 2023

  8. [8]

    Perspective fields for single image camera calibration,

    L. Jin, J. Zhang, Y . Hold-Geoffroy, O. Wang, K. Blackburn-Matzen, M. Sticha, and D. F. Fouhey, “Perspective fields for single image camera calibration,” in CVPR, 2023, pp. 17 307–17 316

    work page 2023

Show all 55 references
  1. [9]

    Tame a wild camera: In-the- wild monocular camera calibration,

    S. Zhu, A. Kumar, M. Hu, and X. Liu, “Tame a wild camera: In-the- wild monocular camera calibration,” in NeurIPS, 2023

    2023

  2. [10]

    CAM-Convs: Camera-Aware Multi-Scale Convolutions for Single-View Depth,

    J. M. Facil, B. Ummenhofer, H. Zhou, L. Montesano, T. Brox, and J. Civera, “CAM-Convs: Camera-Aware Multi-Scale Convolutions for Single-View Depth,” in CVPR, 2019, pp. 11 826–11 835

    2019

  3. [11]

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

    W. Yin, C. Zhang, H. Chen, Z. Cai, G. Yu, K. Wang, X. Chen, and C. Shen, “Metric3d: Towards zero-shot metric 3d prediction from a single image,” in ICCV, 2023, pp. 9043–9053

    2023

  4. [12]

    To- wards zero-shot scale-aware monocular depth estimation,

    V . Guizilini, I. Vasiljevic, D. Chen, R. Ambrus ,, and A. Gaidon, “To- wards zero-shot scale-aware monocular depth estimation,” in ICCV, 2023, pp. 9233–9243

    2023

  5. [13]

    Indoor segmen- tation and support inference from rgbd images,

    N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor segmen- tation and support inference from rgbd images,” in ECCV, 2012, pp. 746–760

    2012

  6. [14]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in CVPR, 2012, pp. 3354– 3361

    2012

  7. [15]

    Google street view: Capturing the world at street level,

    D. Anguelov, C. Dulong, D. Filip, C. Frueh, S. Lafon, R. Lyon, A. Ogale, L. Vincent, and J. Weaver, “Google street view: Capturing the world at street level,” Computer, vol. 43, no. 6, pp. 32–38, 2010

    2010

  8. [16]

    Taskonomy: Disentangling task transfer learning,

    A. R. Zamir, A. Sax, W. Shen, L. J. Guibas, J. Malik, and S. Savarese, “Taskonomy: Disentangling task transfer learning,” inCVPR, 2018, pp. 3712–3722

    2018

  9. [17]

    Shape, illumination, and reflectance from shading,

    J. T. Barron and J. Malik, “Shape, illumination, and reflectance from shading,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 37, no. 8, pp. 1670–1687, 2014

    2014

  10. [18]

    Pixel2mesh: Generating 3d mesh models from single rgb images,

    N. Wang, Y . Zhang, Z. Li, Y . Fu, W. Liu, and Y .-G. Jiang, “Pixel2mesh: Generating 3d mesh models from single rgb images,” in ECCV, 2018, pp. 52–67

    2018

  11. [19]

    Learning shape priors for single-view 3d completion and reconstruction,

    J. Wu, C. Zhang, X. Zhang, Z. Zhang, W. T. Freeman, and J. B. Tenenbaum, “Learning shape priors for single-view 3d completion and reconstruction,” in ECCV, 2018, pp. 646–662

    2018

  12. [20]

    Corenet: Coherent 3d scene reconstruction from a single rgb image,

    S. Popov, P. Bauszat, and V . Ferrari, “Corenet: Coherent 3d scene reconstruction from a single rgb image,” in ECCV, 2020, pp. 366– 383

    2020

  13. [21]

    Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization,

    S. Saito, Z. Huang, R. Natsume, S. Morishima, A. Kanazawa, and H. Li, “Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization,” in ICCV, 2019, pp. 2304–2314

    2019

  14. [22]

    Pifuhd: Multi-level pixel- aligned implicit function for high-resolution 3d human digitization,

    S. Saito, T. Simon, J. Saragih, and H. Joo, “Pifuhd: Multi-level pixel- aligned implicit function for high-resolution 3d human digitization,” in CVPR, 2020, pp. 84–93

    2020

  15. [23]

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

    A. Saxena, M. Sun, and A. Y . Ng, “Make3d: Learning 3d scene structure from a single still image,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 31, no. 5, pp. 824–840, 2008

    2008

  16. [24]

    Learning to recover 3d scene shape from a single image,

    W. Yin, J. Zhang, O. Wang, S. Niklaus, L. Mai, S. Chen, and C. Shen, “Learning to recover 3d scene shape from a single image,” in CVPR, 2021, pp. 204–213

    2021

  17. [25]

    Single- stage 3d geometry-preserving depth estimation model training on dataset mixtures with uncalibrated stereo data,

    N. Patakin, A. V orontsova, M. Artemyev, and A. Konushin, “Single- stage 3d geometry-preserving depth estimation model training on dataset mixtures with uncalibrated stereo data,” in CVPR, 2022, pp. 1705–1714

    2022

  18. [26]

    Depth map prediction from a single image using a multi-scale deep network,

    D. Eigen, C. Puhrsch, and R. Fergus, “Depth map prediction from a single image using a multi-scale deep network,” in NeurIPS, 2014

    2014

  19. [27]

    Va-depthnet: A variational approach to single image depth prediction,

    C. Liu, S. Kumar, S. Gu, R. Timofte, and L. Van Gool, “Va-depthnet: A variational approach to single image depth prediction,” arXiv preprint arXiv:2302.06556, 2023

    2023 arXiv

  20. [28]

    Binsformer: Revisit- ing adaptive bins for monocular depth estimation,

    Z. Li, X. Wang, X. Liu, and J. Jiang, “Binsformer: Revisit- ing adaptive bins for monocular depth estimation,” arXiv preprint arXiv:2204.00987, 2022

    2022 arXiv

  21. [29]

    Localbins: Improving depth estimation by learning local distributions,

    S. F. Bhat, I. Alhashim, and P. Wonka, “Localbins: Improving depth estimation by learning local distributions,” in ECCV, 2022, pp. 480– 496

    2022

  22. [30]

    Transformer-based attention networks for continuous pixel-wise prediction,

    G. Yang, H. Tang, M. Ding, N. Sebe, and E. Ricci, “Transformer-based attention networks for continuous pixel-wise prediction,” in ICCV, 2021, pp. 16 269–16 279

    2021

  23. [31]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” in ICCV, 2021, pp. 12 179–12 188

    2021

  24. [32]

    Zoedepth: Zero-shot transfer by combining relative and metric depth,

    S. F. Bhat, R. Birkl, D. Wofk, P. Wonka, and M. M ¨uller, “Zoedepth: Zero-shot transfer by combining relative and metric depth,” arXiv preprint arXiv:2302.12288, 2023

    2023 arXiv

  25. [33]

    Towards zero-shot scale-aware monocular depth estimation,

    V . Guizilini, I. Vasiljevic, D. Chen, R. Ambrus ,, and A. Gaidon, “Towards zero-shot scale-aware monocular depth estimation,” in Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 9233–9243

    2023

  26. [34]

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

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” in CVPR, 2024, pp. 10 371–10 381

    2024

  27. [35]

    A flexible new technique for camera calibration,

    Z. Zhang, “A flexible new technique for camera calibration,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 22, no. 11, pp. 1330–1334, 2000

    2000

  28. [36]

    Camera calibration with one-dimensional objects,

    ——, “Camera calibration with one-dimensional objects,” IEEE Trans- actions on Pattern Analysis and Machine Intelligence , vol. 26, no. 7, pp. 892–899, 2004

    2004

  29. [37]

    Atlanta world: An expectation max- imization framework for simultaneous low-level edge grouping and camera calibration in complex man-made environments,

    G. Schindler and F. Dellaert, “Atlanta world: An expectation max- imization framework for simultaneous low-level edge grouping and camera calibration in complex man-made environments,” in CVPR, vol. 1, 2004, pp. I–I

    2004

  30. [38]

    A minimum error vanishing point detection approach for uncalibrated monocular images of man-made environments,

    Y . Xu, S. Oh, and A. Hoogs, “A minimum error vanishing point detection approach for uncalibrated monocular images of man-made environments,” in CVPR, 2013, pp. 1376–1383

    2013

  31. [39]

    Robust camera self-calibration from monocular images of manhattan worlds,

    H. Wildenauer and A. Hanbury, “Robust camera self-calibration from monocular images of manhattan worlds,” in CVPR, 2012, pp. 2831– 2838

    2012

  32. [40]

    Automatic camera calibration from a single manhattan image,

    J. Deutscher, M. Isard, and J. MacCormick, “Automatic camera calibration from a single manhattan image,” in ECCV, 2002, pp. 175– 188

    2002

  33. [41]

    Manhattan world: Compass direc- tion from a single image by bayesian inference,

    J. M. Coughlan and A. L. Yuille, “Manhattan world: Compass direc- tion from a single image by bayesian inference,” in ICCV, 1999, pp. 941–947

    1999

  34. [42]

    Lsd: A fast line segment detector with a false detection control,

    R. G. V on Gioi, J. Jakubowicz, J.-M. Morel, and G. Randall, “Lsd: A fast line segment detector with a false detection control,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 32, no. 4, pp. 722–732, 2008

    2008

  35. [43]

    Edlines: A real-time line segment detector with a false detection control,

    C. Akinlar and C. Topal, “Edlines: A real-time line segment detector with a false detection control,” Pattern Recognition Letters , vol. 32, no. 13, pp. 1633–1642, 2011

    2011

  36. [44]

    A perceptual measure for deep single image camera calibration,

    Y . Hold-Geoffroy, K. Sunkavalli, J. Eisenmann, M. Fisher, E. Gam- baretto, S. Hadap, and J.-F. Lalonde, “A perceptual measure for deep single image camera calibration,” in CVPR, 2018, pp. 2354–2363

    2018

  37. [45]

    Ctrl-c: Camera calibration transformer with line-classification,

    J. Lee, H. Go, H. Lee, S. Cho, M. Sung, and J. Kim, “Ctrl-c: Camera calibration transformer with line-classification,” in ICCV, 2021, pp. 16 228–16 237

    2021

  38. [46]

    Neural geometric parser for single image camera calibration,

    J. Lee, M. Sung, H. Lee, and J. Kim, “Neural geometric parser for single image camera calibration,” in ECCV, 2020, pp. 541–557

    2020

  39. [47]

    UniDepth: Universal monocular metric depth estimation,

    L. Piccinelli, Y .-H. Yang, C. Sakaridis, M. Segu, S. Li, L. Van Gool, and F. Yu, “UniDepth: Universal monocular metric depth estimation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2024, pp. 10 106–10 116

    2024

  40. [48]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

    2016

  41. [49]

    Hierarchical chamfer matching: A parametric edge matching algorithm,

    G. Borgefors, “Hierarchical chamfer matching: A parametric edge matching algorithm,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 10, no. 6, pp. 849–865, 1988

    1988

  42. [50]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in ICCV, 2021, pp. 10 012–10 022

    2021

  43. [51]

    Sun rgb-d: A rgb-d scene understanding benchmark suite,

    S. Song, S. P. Lichtenberg, and J. Xiao, “Sun rgb-d: A rgb-d scene understanding benchmark suite,” in CVPR, 2015, pp. 567–576

    2015

  44. [52]

    From big to small: Multi-scale local planar guidance for monocular depth estimation,

    J. H. Lee, M.-K. Han, D. W. Ko, and I. H. Suh, “From big to small: Multi-scale local planar guidance for monocular depth estimation,” arXiv preprint arXiv:1907.10326 , 2019

    1907 arXiv

  45. [53]

    Oasis: A large-scale dataset for single image 3d in the wild,

    W. Chen, S. Qian, D. Fan, N. Kojima, M. Hamilton, and J. Deng, “Oasis: A large-scale dataset for single image 3d in the wild,” in CVPR, 2020, pp. 679–688

    2020

  46. [54]

    P3depth: Monoc- ular depth estimation with a piecewise planarity prior,

    V . Patil, C. Sakaridis, A. Liniger, and L. Van Gool, “P3depth: Monoc- ular depth estimation with a piecewise planarity prior,” inCVPR, 2022, pp. 1610–1621

    2022

  47. [55]

    Automatic upright adjustment of photographs with robust camera calibration,

    H. Lee, E. Shechtman, J. Wang, and S. Lee, “Automatic upright adjustment of photographs with robust camera calibration,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 36, no. 5, pp. 833–844, 2013. APPENDIX VII. P ROOF OF PROPOSITION In this study, we ex...

    2013

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.