Pith. sign in

REVIEW 4 major objections 5 minor 87 references

PacGDC: Label-Efficient Generalizable Depth Completion with Projection Ambiguity and Consistency

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read PacGDC claims that randomizing the scale of pseudo depth labels trains depth-completion models to generalize to unseen scenes with minimal or no metric-depth annotation, reporting state-of-the-art zero-shot and few-shot results.

desk verdict Solid label-efficient depth completion via scale-ambiguity augmentation; a missing direct shape-consistency check is the main gap. read the letter →

arxiv 2507.07374 v1 pith:66KGHYBC submitted 2025-07-10 cs.CV

classification cs.CV
keywords depthcompletionzero-shotgeneralizationpseudo-labelsynthesismonocularfoundationmodelsscaleambiguityprojectionconsistencyfew-shotlearningdatadiversity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that generalizable depth completion—predicting dense metric depth in environments a model has never seen—does not require the usual large-scale collections of densely labeled metric depth maps. Its insight is that the same 2D image can be the projection of many different 3D scenes, so training data can be synthesized by deliberately corrupting scene scale: monocular depth foundation models supply dense predictions whose shapes match the image but whose scales are unreliable, and random interpolation, relocation, and unlabeled images multiply those pseudo geometries cheaply. Consistency is enforced by construction, because the sparse depth input is subsampled from the very pseudo label the network is trained to predict, while shape stays tied to image semantics. If the reported results hold, label-efficient or label-free training can match or approach fully supervised depth completion on unseen domains, with zero added inference cost.

What carries the argument

At the center stands a label-generation identity, equation (3): the pseudo dense label is $\hat d = \theta\big(\sum_{t} \lambda_t R_t(I) + (1-\sum_{t} \lambda_t)d\big)$, where $d$ is the ground-truth depth of a labeled image, each $R_t$ is a monocular depth foundation model (the paper uses DepthAnything and DepthPro), $\lambda_t$ are random interpolation weights, and $\theta$ is a random relocation of the resulting depth field. The equation converts the known scale ambiguity of monocular depth estimators—their predictions are shape-correct but globally and locally scale-wrong—into a mechanism for generating many consistent training geometries from one image. Sparse depth maps are then subsampled from $\hat d$ using uniform sampling and LiDAR/VIO-style patterns, which is what makes position consistency automatic: the network is asked to predict a dense label from points that were literally drawn from that label, while shape consistency is inherited from the foundation model's image-aligned prediction.

What would settle it

Take a held-out set with ground-truth depth and compute each foundation model's residual error after removing per-image scale, for example by standardizing prediction and ground truth to the same mean and variance; if standardized residuals are large, the predicted shapes are not actually consistent with the images. A decisive experiment would replace one foundation model with a deliberately shape-inconsistent pseudo-label source, such as random smooth depth fields matched in mean and variance, and observe whether zero-shot RMSE collapses—if it does not, mere scale diversity, not shape consistency, is carrying the gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that the inherent ambiguity of 2D-to-3D projection is not a nuisance but a resource: because the same 2D image can arise from many 3D scenes, a depth-completion network can be trained on a scene under many synthesized pseudo geometries instead of on expensive metric ground truth. The authors decompose this ambiguity into shape (which 3D form an object takes) and position (where and at what scale it sits), and observe that the two inputs of a depth-completion triplet resolve it: the image fixes shape, and the sparse depth points fix position. They then build a synthesis pipeline that turns this observation into data: monocular depth foundation models supply dense predictions that are shape-consistent with the image but wrong in scale, and random interpolation with ground-truth depth, random relocation, and 390K unlabeled images multiply the available geometries for the same visual input. Sparse depth inputs are subsampled from the pseudo dense labels, so every synthesized triplet is internally consistent. The paper reports that this recipe yields the lowest average RMSE (1966 mm) and MAE (731 mm) across six uniform-sampling zero-shot benchmarks, the lowest average RMSE (1147 mm) and MAE (342 mm) on the VOID and KITTI sensor-pattern benchmarks, and few-shot KITTI results in which 1000 training samples reach 830 mm RMSE and even a single labeled sample beats self-supervised full-shot baselines in MAE.

Load-bearing premise

The pipeline rests on the assumption that the dense depth predictions of the foundation models are consistent with the image in shape (not scale), and that this shape consistency survives random interpolation and relocation; the paper supports this only indirectly through ablations, never by measuring shape consistency directly against real geometry.

Editorial extensions

If this is right

  • Zero-shot depth completion reaches state-of-the-art error levels on six uniform-sampling benchmarks and on the VOID/KITTI sensor-pattern benchmarks without any new metric-depth annotation.
  • The synthesized pseudo triplets transfer across very different sparse-input types (uniform sampling, VIO feature points, and 4/8/16/32/64-line LiDAR), so one pretrained model serves multiple sensor geometries.
  • Few-shot fine-tuning on KITTI with as few as one labeled sample beats all self-supervised full-shot baselines in MAE, and with 1000 samples reaches 830 mm RMSE, approaching or exceeding several full-shot supervised methods.
  • Inference speed and cost are unchanged by the synthesis pipeline, because all label generation happens offline before training.
  • Fine-tuning the pretrained weights on the full KITTI set yields in-domain results competitive with recent specialized depth-completion methods even without spatial-propagation modules.

Reading between the lines

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

  • If shape consistency is the active ingredient, the pipeline should improve automatically as better monocular depth models appear: each new model can be added as another $R_t$ in equation (3) and expands the covered geometry without any new annotation; conversely, the method inherits the worst shape errors of the models in its ensemble.
  • A direct stress test the paper does not run: remove the ground-truth term $(1-\sum_t \lambda_t)d$ entirely and rely on relocation for position information; success would show that labeled depth maps are not needed at all for the synthesis, only unlabeled images.
  • The same 'standardize away scale during training, let sparse measurements restore it' logic could transfer to other geometric tasks where a sparse signal regularizes a fundamentally ambiguous dense prediction, such as surface-normal or occupancy estimation, though the paper does not claim this.
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 / 5 minor

Summary. PacGDC proposes a label-efficient training-data synthesis pipeline for depth completion. It uses multiple monocular depth foundation models (DepthAnything, DepthPro) as "scale manipulators" to produce pseudo dense depth labels, then randomly interpolates these with ground-truth depth and applies a random relocation factor (Eq. 3). Sparse depth maps are subsampled from the pseudo labels, and unlabeled images from SA1B are included, yielding a large set of pseudo training triplets without new annotations. The method is evaluated under zero-shot depth completion on six uniform-sampling benchmarks plus VOID and KITTI sensor-pattern benchmarks, and under few-shot KITTI with 1–1000 samples. The paper reports the lowest average RMSE/MAE in most zero-shot settings and large few-shot gains over retrained baselines, with ablations showing each synthesis component contributes.

Significance. If the reported results hold, PacGDC is a meaningful advance in label-efficient depth completion: it shows that synthetic pseudo-labels from depth foundation models can match or exceed fully supervised methods on unseen domains, and that unlabeled images can be exploited without any metric-depth annotation. The paper is empirically extensive (multiple benchmarks, sparse-depth patterns, zero-shot and few-shot protocols) and includes informative ablations (Tables 6 and 12) as well as released code. The conceptual framing of projection ambiguity and consistency is interesting and could inspire further work. However, the central geometric claim of the paper — that interpolation and relocation preserve shape consistency — is asserted but never directly verified, and several experimental details (random-factor distributions, error bars, loss confounds) are left unspecified.

major comments (4)
  1. [§3.3, Eq. (3)] The central assumption that the synthesized pseudo labels preserve shape consistency with the image is never directly tested. The paper claims that interpolation and relocation maintain "projection consistency" and attributes the generalization gains to this property, but all supporting evidence is end-task performance (Table 6). Because the loss in Eq. (4) combines a standardized term and a raw L1 term, and because sparse inputs are subsampled from the very same pseudo dense labels, a model could in principle fit the standardized labels without learning image-consistent 3D geometry, while still achieving low RMSE on benchmarks. Please provide a direct measurement of shape consistency on a labeled subset, for example relative depth error, ordinal inconsistency, or normal-angle error between the synthesized labels and the corresponding ground truth, and show how consistency varies with the interpolation factor λt and relocation factor θ. Without such a measurement, the paper's core theoretical claim is unsupported.
  2. [§4.1, Tables 4–5] The few-shot results are reported without error bars or multiple seeds. With only 1, 10, and 100 training samples, run-to-run variance is likely substantial, and the claimed improvement over baselines could be within noise. Please report mean ± std over at least 3 random seeds for the main few-shot comparisons, or provide a clear justification for why variance is negligible. The current tables give no indication of stability, making it difficult to judge the significance of the reported gains.
  3. [§3.3, Eqs. (2)–(3)] The sampling distributions of the interpolation factors λt and the relocation factor θ are not specified anywhere in the paper or supplement. These random factors are the core of the synthesis pipeline, so the method is not reproducible without them. Please give the exact distributions (ranges, prior probabilities), including how P(Interpolation) is applied in the multi-model case and the distribution of θ (scale factor versus offset, allowed range). This is also needed to assess the effective geometry diversity introduced by the pipeline.
  4. [§3.4, Eq. (4)] The loss function includes a standardized term T and a raw L1 term, but the paper does not isolate the contribution of the standardization to the observed gains. Since the ablation study varies the synthesis components, it is possible that part of the improvement comes from the loss's ability to fit normalized pseudo labels rather than from the shape consistency of the labels themselves. Please report an ablation that trains with and without the standardized term (while keeping the synthesis pipeline fixed), or otherwise show that the gains are not due solely to this loss design.
minor comments (5)
  1. [§4.1, ``Evaluation Protocol''] The sentence ``Following zero-shot depth estimation [69], we impose no restrictions on the model training, evaluating only released models on the same test setups'' is ambiguous; it should clarify that all zero-shot models are trained on the source datasets only and evaluated on held-out target datasets without fine-tuning.
  2. [Table 5, SparseDC row] The reported iRMSE of 12598.67 for SparseDC at 1 shot is many orders of magnitude larger than other values in the same table and is likely a unit or computation artifact. Please check the metric computation and correct the value or add a note about the scaling.
  3. [Supplementary, Sec. 6] In the text accompanying Table 8, the statement that ``Ours-T even outperforms SPNet-L'' is only true for a model trained on 25% of the training data, as noted in the table caption. Please qualify this claim in the main text or remove the comparison to avoid misleading readers.
  4. [§4.4, ``Interpolated vs Original Labels''] The sentence ``The remaining probability is used for randomly selecting original dense depth maps, including ground-truth depth maps and dense predictions from DepthAnything'' is vague: it should specify whether the random selection is uniform, and whether the original dense predictions are pre-computed or generated on the fly.
  5. [Throughout] Several references to equations and supplementary sections are missing or abbreviated (e.g., Eq. (4) is referenced but only appears in the supplement). Please ensure all cited equations are numbered and appear in the main text or supplement consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the synthesis pipeline is self-contained and the reported benchmark results are not forced by construction.

full rationale

PacGDC's claimed derivation chain begins with the pinhole projection identity in Eq. (1), which is an external geometric fact, and then constructs pseudo triplets by applying fixed pretrained monocular depth models with random interpolation and relocation factors (Eqs. (2)-(3)), subsampling sparse depths from the same pseudo labels, and training SPNet with the loss in Eq. (4). No test-benchmark quantity is fitted: the interpolation and relocation factors are random draws, the test sets are held out, and no model selection is made on the reported validation errors. The position-consistency property is true by construction, since each sparse map is subsampled from its corresponding pseudo dense label, while the shape-consistency of foundation-model predictions is an empirical assumption about external models rather than a definitional restatement of the reported RMSE/MAE. The paper self-cites SPNet and G2-MonoDepth for the architecture and base loss, but the central ablation in Tab. 6 compares against the same SPNet baseline and attributes gains to the proposed synthesis components, so these self-citations are not load-bearing in a way that forces the zero-shot results. The absence of a direct geometric consistency measurement for Eq. (3) labels is a verification gap and a correctness risk, but it is not circular. Potential pretraining overlap between the depth foundation models and test scenes is a data-contamination concern, not an in-paper circular reduction, and the text provides no evidence of such overlap. Overall, the derivation is self-contained against the held-out benchmarks.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The core technical contribution is a training-data recipe built on pretrained depth foundation models rather than a closed-form derivation. The central claim rests on the projective identity (Eq. 1), on the empirical shape consistency of foundation-model depth, and on the validity of the standardized loss in Eq. (4). The random interpolation and relocation factors are hyperparameters whose distributions are not specified, so they are counted as free parameters despite being random draws.

free parameters (5)
  • Interpolation factor λt = random, distribution unspecified
    Controls the mixing ratio between each foundation model's pseudo depth and the ground-truth depth in Eq. (3). The paper only states that λt is random with Σλt ≤ 1, without giving the sampling distribution, so the augmentation strength is an unstated choice.
  • Relocation factor θ = random, distribution unspecified
    Randomly relocates the interpolated depth map to a new position/scale in Eq. (3). The range or distribution is not reported, making the geometry diversity non-reproducible from the text.
  • P(Interpolation) = 1.0
    Probability of using interpolated labels vs original dense maps; swept in Tab. 6, final setting uses 1.0. This is a hand-selected hyperparameter.
  • Number of pseudo labels N and sparse maps M per image = not reported
    Defines the multiplicity of synthetic triplets per visual scene in Sec. 3.2; the paper does not state the final N and M, which directly affect training data size and diversity.
  • Number of foundation models L = 2 (main); 4 (ablation Tab. 12)
    The paper adopts DepthAnything and DepthPro in the main results and tries four models in the ablation. This is a design choice influencing diversity and computational cost.
assumptions (5)
  • standard math Pinhole camera projection: scaling a depth value d_i by α maps to a valid 3D point (αx_i, αy_i, αz_i) that projects to the same 2D pixel.
    Used in Eq. (1) to establish projection ambiguity. This is elementary projective geometry and assumed without proof.
  • domain assumption Depth foundation models (DepthAnything, DepthPro) produce dense depth predictions whose shape is consistent with the image semantics, although their scene scales are inaccurate.
    Sec. 3.3 relies on this to guarantee the synthesized pseudo labels retain shape consistency. The paper cites the models' robustness but does not measure shape consistency directly.
  • domain assumption Random interpolation between ground-truth depth and pseudo depth, followed by random relocation (Eq. 3), preserves the shape cues needed for training.
    There is no mathematical guarantee that an affine combination of two shape-consistent depth maps remains shape-consistent with the image; the paper supports this only through ablation results (Tab. 6).
  • domain assumption Sparse depth points subsampled from dense pseudo labels (uniform sampling via [51], LiDAR/SFM patterns via [82]) are representative of real sensor sparse depths.
    Sec. 3.2 adopts prior subsampling pipelines as valid proxies for actual sensor patterns; no new validation is provided for the synthetic settings.
  • domain assumption The loss in Eq. (4), which includes a mean-deviation-standardized depth term T from G2-MonoDepth [50], is an appropriate objective for learning generalizable metric depth completion from scale-ambiguous pseudo labels.
    The standardized term makes the loss partially scale-invariant, which is what allows scale-ambiguous labels to teach shape; the paper does not ablate the loss choice separately from the synthetic data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PacGDC: Label-Efficient Generalizable Depth Completion with Projection Ambiguity and Consistency." pith.science (2026). https://pith.science/paper/66KGHYBC

@misc{pith2026250707374,
  author       = {Pith},
  title        = {Pith review of: PacGDC: Label-Efficient Generalizable Depth Completion with Projection Ambiguity and Consistency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66KGHYBC}},
  note         = {Machine review of arXiv:2507.07374}
}
read the original abstract

Generalizable depth completion enables the acquisition of dense metric depth maps for unseen environments, offering robust perception capabilities for various downstream tasks. However, training such models typically requires large-scale datasets with metric depth labels, which are often labor-intensive to collect. This paper presents PacGDC, a label-efficient technique that enhances data diversity with minimal annotation effort for generalizable depth completion. PacGDC builds on novel insights into inherent ambiguities and consistencies in object shapes and positions during 2D-to-3D projection, allowing the synthesis of numerous pseudo geometries for the same visual scene. This process greatly broadens available geometries by manipulating scene scales of the corresponding depth maps. To leverage this property, we propose a new data synthesis pipeline that uses multiple depth foundation models as scale manipulators. These models robustly provide pseudo depth labels with varied scene scales, affecting both local objects and global layouts, while ensuring projection consistency that supports generalization. To further diversify geometries, we incorporate interpolation and relocation strategies, as well as unlabeled images, extending the data coverage beyond the individual use of foundation models. Extensive experiments show that PacGDC achieves remarkable generalizability across multiple benchmarks, excelling in diverse scene semantics/scales and depth sparsity/patterns under both zero-shot and few-shot settings. Code: https://github.com/Wang-xjtu/PacGDC.

Figures

Figures reproduced from arXiv: 2507.07374 by the authors.

Figure 1
Figure 1. PacGDC generalizes effectively across unseen scenarios with a wide range of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the ambiguity and consistency in 2D-to-3D projections for generalizable depth completion. (a) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Data distribution of our synthesis method on 1000 samples from UnrealCV dataset [ [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Overview of the proposed data synthesis pipeline, which leverages multiple depth foundation models, interpolation and relocation [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Illustration of basic geometry synthesis. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: “Few-shot vs full-shot” on KITTI with 64/32/16/8/4 lines LiDAR. Our model is trained with 1000 samples, while the baselines use 86K samples including SparseDC [25], PackNet￾SAN [13], PeNet [15], and SPAgNet [8]. model, we further evaluate it, using less than 1000 sampl…
Figure 8
Figure 8. Figure 8: Zero-shot depth completion on unseen scenarios with different scene semantics/scales and depth sparsity/patterns [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Few-shot depth completion on KITTI with 8- and 16-lines LiDAR, using models trained with 1 and 10 samples, respectively [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Few-shot depth completion on KITTI with 32- and 64-lines LiDAR, using models trained with 100 and 1000 samples, respectively [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

87 extracted references · 63 canonical work pages

  1. [1]

    Revisiting depth completion from a stereo matching perspective for cross-domain generalization

    Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In 2024 International Conference on 3D Vision (3DV), pages 1360–1370. IEEE, 2024. 1

  2. [2]

    Depth pro: Sharp monocular metric depth in less than a second

    Aleksei Bochkovskii, Ama ¨el 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 preprint arXiv:2410.02073, 2024. 2, 3, 5, 6, 8

  3. [3]

    A naturalistic open source movie for op- tical flow evaluation

    Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for op- tical flow evaluation. In European Conference on Computer Vision, pages 611–625. Springer, 2012. 1, 6

  4. [4]

    Orb-slam3: An accu- rate open-source library for visual, visual–inertial, and mul- timap slam

    Carlos Campos, Richard Elvira, Juan J G ´omez Rodr´ıguez, Jos´e MM Montiel, and Juan D Tard´os. Orb-slam3: An accu- rate open-source library for visual, visual–inertial, and mul- timap slam. IEEE Transactions on Robotics , 37(6):1874– 1890, 2021. 1

  5. [5]

    Matterport3d: Learning from rgb-d data in indoor environments

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 5

  6. [6]

    Depth com- pletion using geometry-aware embedding

    Hu Chen, Hongyu Yang, Yi Zhang, et al. Depth com- pletion using geometry-aware embedding. In 2022 Inter- national Conference on Robotics and Automation (ICRA) , pages 8680–8686. IEEE, 2022. 7

  7. [7]

    Learning depth with convolutional spatial propagation network

    Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE transactions on pattern analysis and machine intelligence , 42(10):2361–2379, 2019. 1, 3

  8. [8]

    Spar- sity agnostic depth completion

    Andrea Conti, Matteo Poggi, and Stefano Mattoccia. Spar- sity agnostic depth completion. In Proceedings of the ieee/cvf winter conference on applications of computer vi- sion, pages 5871–5880, 2023. 7

Show all 87 references
  1. [9]

    Deep ordinal regression net- work for monocular depth estimation

    Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat- manghelich, and Dacheng Tao. Deep ordinal regression net- work for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 2002–2011, 2018. 3

  2. [10]

    Virtual worlds as proxy for multi-object tracking anal- ysis

    Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking anal- ysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340–4349, 2016. 5

  3. [11]

    Are we ready for autonomous driving? the kitti vision bench- mark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision bench- mark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3354–

  4. [12]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485–2494, 2020. 1

  5. [13]

    Sparse auxiliary networks for unified monocular depth prediction and completion

    Vitor Guizilini, Rares Ambrus, Wolfram Burgard, and Adrien Gaidon. Sparse auxiliary networks for unified monocular depth prediction and completion. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition, pages 11078–11088, 2021. 7

  6. [14]

    Distill any depth: Distillation cre- ates a stronger monocular depth estimator

    Xiankang He, Dongyan Guo, Hongji Li, Ruibo Li, Ying Cui, and Chi Zhang. Distill any depth: Distillation cre- ates a stronger monocular depth estimator. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 2

  7. [15]

    Penet: Towards precise and efficient image guided depth completion

    Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In 2021 IEEE International Con- ference on Robotics and Automation (ICRA) , pages 13656– 13662. IEEE, 2021. 7

  8. [16]

    Depth com- pletion with twin surface extrapolation at occlusion bound- aries

    Saif Imran, Xiaoming Liu, and Daniel Morris. Depth com- pletion with twin surface extrapolation at occlusion bound- aries. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 2583–2592,

  9. [17]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9492...

  10. [18]

    Con- match: Semi-supervised learning with confidence-guided consistency regularization

    Jiwon Kim, Youngjo Min, Daehwan Kim, Gyuseong Lee, Junyoung Seo, Kwangrok Ryoo, and Seungryong Kim. Con- match: Semi-supervised learning with confidence-guided consistency regularization. In European Conference on Computer Vision, pages 674–690. Springer, 2022. 3

  11. [19]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 202...

  12. [20]

    Evaluation of cnn-based single-image depth estimation methods

    Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Con- ference on Computer Vision (ECCV) Workshops, pages 0–0,

  13. [21]

    Comatch: Semi-supervised learning with contrastive graph regulariza- tion

    Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regulariza- tion. In Proceedings of the IEEE/CVF international confer- ence on computer vision, pages 9475–9484, 2021. 3

  14. [22]

    Semireward: A general reward model for semi-supervised learning

    Siyuan Li, Weiyang Jin, Zedong Wang, Fang Wu, Zicheng Liu, Cheng Tan, and Stan Z Li. Semireward: A general reward model for semi-supervised learning. arXiv preprint arXiv:2310.03013, 2023. 3

  15. [23]

    Dis- tilling monocular foundation model for fine-grained depth completion

    Yingping Liang, Yutao Hu, Wenqi Shao, and Ying Fu. Dis- tilling monocular foundation model for fine-grained depth completion. In Proceedings of the Computer Vision and Pat- tern Recognition Conference, pages 22254–22265, 2025. 1

  16. [24]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 1

  17. [25]

    Sparsedc: Depth completion from sparse and non-uniform inputs

    Chen Long, Wenxiao Zhang, Zhe Chen, Haiping Wang, Yuan Liu, Peiling Tong, Zhen Cao, Zhen Dong, and Bisheng Yang. Sparsedc: Depth completion from sparse and non-uniform inputs. Information Fusion, 110:102470, 2024. 3, 7

  18. [26]

    Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera

    Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In 2019 International Conference on Robotics and Automation (ICRA), pages 3288–3295. IEEE, 2019. 7

  19. [27]

    Occdepth: A depth-aware method for 3d semantic scene completion

    Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1

  20. [28]

    Semattnet: Toward attention- based semantic aware guided depth completion

    Danish Nazir, Alain Pagani, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. Semattnet: Toward attention- based semantic aware guided depth completion. IEEE Ac- cess, 10:120781–120791, 2022. 3

  21. [29]

    Non-local spatial propagation network for depth completion

    Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In European Conference on Computer Vision , pages 120–136. Springer, 2020. 3, 6, 7, 1

  22. [30]

    A simple yet universal framework for depth completion

    Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In The Thirty-eighth An- nual Conference on Neural Information Processing Systems,

  23. [31]

    Learning affinity with hyperbolic representation for spatial propagation

    Jin-Hwi Park, Jaesung Choe, Inhwan Bae, and Hae-Gon Jeon. Learning affinity with hyperbolic representation for spatial propagation. In Proceedings of the International Con- ference on Machine Learning (ICML), 2023. 3

  24. [32]

    Depth prompting for sensor-agnostic depth estimation

    Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9859–9869, 2024. 2

  25. [33]

    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. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106–10116, 2024. 3, 5

  26. [34]

    Deepli- dar: Deep surface normal guided depth prediction for out- door scene from sparse lidar data and single color image

    Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deepli- dar: Deep surface normal guided depth prediction for out- door scene from sparse lidar data and single color image. In Proceedings of the IEEE/CVF Conference on Comput...

  27. [35]

    Few- shot depth completion using denoising diffusion probabilis- tic model

    Weihang Ran, Wei Yuan, and Ryosuke Shibasaki. Few- shot depth completion using denoising diffusion probabilis- tic model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6559– 6567, 2023. 3, 2

  28. [36]

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

    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, 3, 4, 5

  29. [37]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. 3

  30. [38]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3

  31. [39]

    Guide- former: Transformers for image guided depth completion

    Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guide- former: Transformers for image guided depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6250–6259, 2022. 3

  32. [40]

    A multi-view stereo benchmark with high- resolution images and multi-camera videos

    Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and An- dreas Geiger. A multi-view stereo benchmark with high- resolution images and multi-camera videos. In Proceed- ings of the IEEE conference on computer vision and ...

  33. [41]

    Towards comprehensive monocular depth estimation: Mul- tiple heads are better than one

    Shuwei Shao, Ran Li, Zhongcai Pei, Zhong Liu, Weihai Chen, Wentao Zhu, Xingming Wu, and Baochang Zhang. Towards comprehensive monocular depth estimation: Mul- tiple heads are better than one. IEEE Transactions on Multi- media, 25:7660–7671, 2022. 3

  34. [42]

    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, pages 746–760. Springer, 2012. 3, 6, 1

  35. [43]

    Fixmatch: Simpli- fying semi-supervised learning with consistency and confi- dence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simpli- fying semi-supervised learning with consistency and confi- dence. Advances in Neural Information Processing Systems, ...

  36. [44]

    Self-supervised depth completion from direct visual-lidar odometry in autonomous driving

    Zhenbo Song, Jianfeng Lu, Yazhou Yao, and Jian Zhang. Self-supervised depth completion from direct visual-lidar odometry in autonomous driving. IEEE Transactions on In- telligent Transportation Systems, 23(8):11654–11665, 2021. 7

  37. [45]

    Bilateral propagation network for depth completion

    Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9763–9772, 2024. 3, 1

  38. [46]

    Diode: A dense indoor and outdoor depth dataset

    Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Moham- madreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 6, 1

  39. [47]

    Marigold-dc: Zero-shot monocular depth completion with guided diffusion

    Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. arXiv preprint arXiv:2412.13389, 2024. 1

  40. [48]

    Depth map recovery based on a unified depth boundary distortion model

    Haotian Wang, Meng Yang, Xuguang Lan, Ce Zhu, and Nan- ning Zheng. Depth map recovery based on a unified depth boundary distortion model. IEEE transactions on image pro- cessing, 31:7020–7035, 2022. 1

  41. [49]

    Rgb-guided depth map recovery by two-stage coarse-to-fine dense crf models

    Haotian Wang, Meng Yang, Ce Zhu, and Nanning Zheng. Rgb-guided depth map recovery by two-stage coarse-to-fine dense crf models. IEEE Transactions on Image Processing, 32:1315–1328, 2023. 2, 1

  42. [50]

    G2- monodepth: A general framework of generalized depth in- ference from monocular rgb+ x data

    Haotian Wang, Meng Yang, and Nanning Zheng. G2- monodepth: A general framework of generalized depth in- ference from monocular rgb+ x data. IEEE Transactions on Pattern Analysis and Machine Intelligence , 46(5):3753– 3771, 2024. 1, 3, 5, 6, 7, 8

  43. [51]

    Scale propagation network for generalizable depth comple- tion

    Haotian Wang, Meng Yang, Xinhu Zheng, and Gang Hua. Scale propagation network for generalizable depth comple- tion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1908–1922, 2025. 2, 3, 4, 5, 6, 7, 8, 1

  44. [52]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697– 20709, 2024. 4

  45. [53]

    Freematch: Self-adaptive thresholding for semi-supervised learning

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 3

  46. [54]

    Lrru: Long-short range recurrent updating networks for depth completion

    Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9422–9432,

  47. [55]

    Improving depth completion via depth feature upsampling

    Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 21104–21113, 2024. 3, 5, 6, 7, 1

  48. [56]

    Monocular visual-inertial depth estimation

    Diana Wofk, Ren ´e Ranftl, Matthias M ¨uller, and Vladlen Koltun. Monocular visual-inertial depth estimation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 6095–6101. IEEE, 2023. 2, 5

  49. [57]

    Unsupervised depth comple- tion with calibrated backprojection layers

    Alex Wong and Stefano Soatto. Unsupervised depth comple- tion with calibrated backprojection layers. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 12747–12756, 2021. 7

  50. [58]

    Unsupervised depth completion from visual iner- tial odometry

    Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual iner- tial odometry. IEEE Robotics and Automation Letters, 5(2): 1899–1906, 2020. 1, 6, 8

  51. [59]

    Con- vnext v2: Co-designing and scaling convnets with masked autoencoders

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Con- vnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133– 161...

  52. [60]

    Convolution with even-sized kernels and symmetric padding

    Shuang Wu, Guanrui Wang, Pei Tang, Feng Chen, and Lup- ing Shi. Convolution with even-sized kernels and symmetric padding. Advances in Neural Information Processing Sys- tems, 32, 2019. 1

  53. [61]

    Augundo: Scaling up augmentations for monocular depth completion and estima- tion

    Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estima- tion. In European Conference on Computer Vision , 2024. 7

  54. [62]

    Structure-guided ranking loss for single im- age depth prediction

    Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single im- age depth prediction. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 611–620, 2020. 3, 5

  55. [63]

    Unsupervised data augmentation for consistency training

    Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Sys- tems, 33:6256–6268, 2020. 3

  56. [64]

    Self-training with noisy student improves imagenet clas- sification

    Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet clas- sification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 10687– 10698, 2020. 3

  57. [65]

    Towards domain-agnostic depth completion

    Guangkai Xu, Wei Yin, Jianming Zhang, Oliver Wang, Si- mon Niklaus, Simon Chen, and Jia-Wang Bian. Towards domain-agnostic depth completion. Machine Intelligence Research, pages 1–18, 2024. 2, 3

  58. [66]

    Depth completion from sparse li- dar data with depth-normal constraints

    Yan Xu, Xinge Zhu, Jianping Shi, Guofeng Zhang, Hujun Bao, and Hongsheng Li. Depth completion from sparse li- dar data with depth-normal constraints. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 2811–2820, 2019. 3

  59. [67]

    Desnet: Decomposed scale-consistent net- work for unsupervised depth completion

    Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent net- work for unsupervised depth completion. In Proceedings of the AAAI conference on artificial intelligence , pages 3109– 3117, 2023. 7

  60. [68]

    Tri- perspective view decomposition for geometry-aware depth completion

    Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri- perspective view decomposition for geometry-aware depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4874– 4884,...

  61. [69]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 2, 3, 5, 6, 8

  62. [70]

    Depth any- thing v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2. Advances in Neural Information Processing Sys- tems, 37:21875–21911, 2024. 2

  63. [71]

    Blendedmvs: A large- scale dataset for generalized multi-view stereo networks

    Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large- scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1790–1799, 2020. 5

  64. [72]

    Virtual normal: En- forcing geometric constraints for accurate and robust depth prediction

    Wei Yin, Yifan Liu, and Chunhua Shen. Virtual normal: En- forcing geometric constraints for accurate and robust depth prediction. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(10):7282–7295, 2021. 2, 4, 5

  65. [73]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043–9053, 2023. 3

  66. [74]

    Neural window fully-connected crfs for monocu- lar depth estimation

    Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocu- lar depth estimation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 3916–3925, 2022. 3

  67. [75]

    Flexmatch: Boosting semi-supervised learning with curricu- lum pseudo labeling

    Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jin- dong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curricu- lum pseudo labeling. Advances in Neural Information Pro- cessing Systems, 34:18408–18419, 2021. 3

  68. [76]

    Multitask gans for se- mantic segmentation and depth completion with cycle con- sistency

    Chongzhen Zhang, Yang Tang, Chaoqiang Zhao, Qiyu Sun, Zhencheng Ye, and J ¨urgen Kurths. Multitask gans for se- mantic segmentation and depth completion with cycle con- sistency. IEEE Transactions on Neural Networks and Learn- ing Systems, 32(12):5404–5415, 2021. 3

  69. [77]

    Deep depth comple- tion of a single rgb-d image

    Yinda Zhang and Thomas Funkhouser. Deep depth comple- tion of a single rgb-d image. InProceedings of the IEEE con- ference on computer vision and pattern recognition , pages 175–185, 2018. 3

  70. [78]

    Completionformer: Depth completion with convolutions and vision transform- ers

    Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transform- ers. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 18527–18536,

  71. [79]

    Bev@ dc: Bird’s-eye view assisted training for depth completion

    Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird’s-eye view assisted training for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233–9242, 2023. 3, 2

  72. [80]

    Semi-supervised learning by disagreement

    Zhi-Hua Zhou and Ming Li. Semi-supervised learning by disagreement. Knowledge and Information Systems, 24:415– 439, 2010. 3

  73. [81]

    Ogni-dc: Robust depth comple- tion with optimization-guided neural iterations

    Yiming Zuo and Jia Deng. Ogni-dc: Robust depth comple- tion with optimization-guided neural iterations. In European Conference on Computer Vision , pages 78–95. Springer,

  74. [82]

    Omni- dc: Highly robust depth completion with multiresolution depth integration

    Yiming Zuo, Willow Yang, Zeyu Ma, and Jia Deng. Omni- dc: Highly robust depth completion with multiresolution depth integration. arXiv preprint arXiv:2411.19278 , 2024. 2, 3, 4, 6, 7, 1 PacGDC: Label-Efficient Generalizable Depth Completion with Projection Ambiguity and Consis...

  75. [84]

    Ours-T” even outperforms “SPNet-L

    More Implementation Details Training Details. Zero-shot Depth Completion: The train- ing data simply concentrate all available training datasets following [50, 51], without any explicit balancing strategies as in [82]. Due to resource constraints, the training res- olution is ...

  76. [85]

    We fur- ther evaluate PacGDC on the DDAD [12] dataset, compar- ing to more generalizable and supervised baselines, follow- ing the standard protocol of VPP4DC [1]

    More Quantitative Results Zero-shot Depth Completion on DDAD Dataset. We fur- ther evaluate PacGDC on the DDAD [12] dataset, compar- ing to more generalizable and supervised baselines, follow- ing the standard protocol of VPP4DC [1]. The baseline results are directly taken fro...

  77. [86]

    We evaluate our approach with four different depth foundation mod- els: DepthAnything (DA) [69], DepthPro [2], DepthAny- thingV2 (DA V2) [70], and DistillAnyDepth (DistillAD) [14]

    More Ablation Study Different Depth Foundation Models. We evaluate our approach with four different depth foundation mod- els: DepthAnything (DA) [69], DepthPro [2], DepthAny- thingV2 (DA V2) [70], and DistillAnyDepth (DistillAD) [14]. As shown in Tab. 12, PacGDC consistently ...

  78. [87]

    We further provide visual examples of zero-shot scenarios in Fig

    More Visual Results Zero-Shot Depth Completion. We further provide visual examples of zero-shot scenarios in Fig. 8, covering a range DA [69] DepthPro [2] DA V2 [70] DistillAD [14]RMSE↓ MAE↓ 2484 990 ✓ 2277 857 ✓ ✓ 2241 854 ✓ ✓ 2243 852 ✓ ✓ 2276 859 ✓ ✓ ✓ 2232 848 ✓ ✓ ✓ ✓ 2279...

  79. [3361]

    1, 2, 3, 6, 8

    IEEE, 2012. 1, 2, 3, 6, 8

Pith tools

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