Pith. sign in

REVIEW 5 major objections 6 minor 73 references

Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Aligning a pre-trained affine-invariant depth diffusion prior with sparse metric measurements at test time yields zero-shot depth completion that generalizes across indoor and outdoor scenes and outperforms test-time adaptation methods by…

desk verdict Solid zero-shot depth completion recipe with broad experiments; the compatibility proof is weak but the empirical results on four domains carry the paper. read the letter →

arxiv 2502.06338 v1 pith:XM2PTG76 submitted 2025-02-10 cs.CV

Lee Hyoseok , Kyeong Seon Kim , Kwon Byung-Ki , Tae-Hyun Oh This is my paper · ORCID
classification cs.CV MSC 68T4568T07
keywords depthcompletionzero-shottest-timealignmentdiffusionprioraffine-invariantdomaingeneralizationoutlierfiltering
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 aims to show that dense metric depth maps can be completed from sparse sensor points without any training on depth-completion datasets. The proposed method combines a pre-trained monocular depth diffusion model, which produces relative (affine-invariant) depth, with an optimization loop at test time that forces the output to match the sparse metric measurements exactly. This zero-shot approach generalizes across indoor and outdoor domains, improving RMSE by up to 21% over test-time adaptation baselines while preserving scene details. If correct, it means the scale and shift ambiguity of generative depth priors can be resolved on the fly using sparse sensors, eliminating the need for domain-specific training data.

What carries the argument

The central object is the reverse diffusion trajectory of a latent diffusion model pre-trained for monocular depth estimation. The key mechanism is test-time alignment: an inner-loop optimization (Eq. 7) over the predicted clean latent $z_0(z_t)$ minimizes an L1 loss against the sparse metric measurements, then re-noises the result to the current timestep via the forward process (Eq. 8) so the trajectory stays on the prior's manifold while respecting hard constraints. This correction step, repeated every few sampling steps, converts an affine-invariant depth prior into a metric depth completion engine, supported by an R-SSIM loss for structure preservation and a superpixel+RANSAC outlier filter for unreliable sensor points.

What would settle it

Run the method on a dataset with physically extreme metric depth ranges (e.g., aerial or satellite scenes with depths from 0.5 m to 5000 m, or a panoramic sensor with non-pinhole projection) where the sparse points are reliable, and compare against high-fidelity ground truth. If the affine-invariant prior's manifold cannot represent these depths, RMSE will degrade sharply and the optimization loop will either diverge or settle on a plausible-looking but wrong structure; a systematic floor in the residual $\|y - A(D(z_0))\|$ after convergence, much larger than sensor noise, would indicate manifold mismatch.

Watch

Extended reading notes

Core claim

A pre-trained monocular depth diffusion model operating in the affine-invariant depth space already represents a rich prior over real-world scene geometry, and this prior can be steered at test time to satisfy sparse metric measurements. The paper demonstrates this through an empirical reconstruction experiment showing that normalized metric depth maps lie near the prior's manifold, and through a complete method that aligns the diffusion sampling trajectory with sparse depth as hard constraints. The method achieves strong zero-shot depth completion across NYUv2, SceneNet, Waymo, nuScenes, KITTI, and VOID without any training on the target domain, outperforming test-time adaptation methods and matching or exceeding several few-shot depth-prior approaches.

Load-bearing premise

The load-bearing premise is that the data manifold learned by the pre-trained relative-depth diffusion model is close enough to the normalized metric depth space that the test-time optimization loop can pull the trajectory onto a metric-consistent solution; the paper supports this only with a limited reconstruction experiment on two completion models, not with a formal characterization.

Editorial extensions

If this is right

  • Depth completion becomes solvable zero-shot for any RGB plus sparse-depth pair, as long as a strong relative-depth prior is available.
  • Affine-invariant depth priors are sufficient for metric depth completion when aligned to sparse metric measurements at test time.
  • The proposed prior-based outlier filtering improves reliability, especially for outdoor LiDAR data with see-through points on glass or other reflective surfaces.
  • The method is compatible with both a 50-step diffusion model and a 1–2 step flow-matching model, with inference time scaling with base model steps (from about 101 s down to 16 s).

Reading between the lines

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

  • The same test-time alignment principle could extend to other inverse problems where a generative prior is affine-invariant, such as relative surface normals, relative lighting, or intrinsic image decomposition, turning sparse absolute measurements into hard constraints.
  • A clear stress test emerges: on scenes whose metric depth distribution falls far outside the prior's training range (extreme distances, non-pinhole projections, unusual sensor noise), the optimization loop should either diverge or settle on a plausible-looking but wrong structure, with the residual $\|y - A(D(z_0))\|$ staying at a floor well above sensor noise.
  • The paper's R-SSIM loss suggests that a deeper learned feature-similarity regularizer might replace the hand-tuned structure term, potentially sharpening edges further without adjusting per-dataset weights.
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

5 major / 6 minor

Summary. The paper proposes a zero-shot depth completion method that combines a pretrained affine-invariant depth diffusion prior (Marigold or DepthFM) with per-sample test-time optimization. Sparse metric depth measurements are enforced through an optimization loop on the denoised latent z0(zt), followed by a noise-remapping step, together with a superpixel/RANSAC outlier filter and an R-SSIM structural regularizer. Experiments on indoor (NYUv2, SceneNet) and outdoor (Waymo, nuScenes, KITTI, VOID) datasets compare favorably against test-time adaptation baselines and unsupervised methods, and the paper reports up to a 21% average performance improvement over prior state-of-the-art test-time adaptation.

Significance. If the central claim holds, the paper offers a useful training-free recipe for leveraging monocular depth foundation models in metric depth completion across domains, and the test-time alignment idea is interesting beyond the specific depth-completion setting. The paper is unusually broad in its evaluation, includes ablations of the main components, and makes an honest attempt to analyze ground-truth sensitivity and stochasticity of diffusion sampling. However, the significance is conditional on resolving several load-bearing issues, in particular the mismatch between the reported nuScenes benchmark and the supplementary description, the unverified compatibility of the affine-invariant prior with random-noise test-time trajectories, and the finite-optimization approximation of the claimed 'hard constraint'.

major comments (5)
  1. [Section 3.2, Eqs. (7)-(9)] The 'hard constraint' characterization is not supported by the described procedure. The optimization loop in Eq. (7) is run for a finite number of iterations (200 per Supplementary A.1), and the remapping in Eq. (8) adds Gaussian noise, so the resulting latent no longer satisfies the sparse-depth consistency that Eq. (7) optimizes. Eq. (9) is then an approximate gradient step rather than a projection onto the measurement set. The paper should either report the residual sparse-depth error after the full sampling loop, or rename the mechanism (e.g., 'strong test-time guidance') and justify why the residual is negligible for the downstream claims.
  2. [Section 3.2 / Eq. (10) and Supplementary Table S2] The central compatibility claim that the affine-invariant depth prior can represent normalized metric depth space is supported only by a reconstruction experiment that starts from a complete depth map, encodes it, adds modest noise, and reverses the diffusion process. This does not test the actual test-time scenario, which begins from pure random noise and relies on sparse guidance to pull the trajectory onto a metric-consistent solution. The supplementary table itself shows severe degradation at timestamp 1000 (RMSE 3.443 to 3.475), the regime closest to the random-noise start. The authors should add a direct test: run the full reverse chain from pure noise with sparse guidance and measure whether the resulting samples preserve metric consistency and lie on the prior manifold, or provide quantitative evidence about the manifold coverage of the optimized latents.
  3. [Section 4.1 and Supplementary B.1] There is a direct contradiction about which ground-truth definition is used for the nuScenes benchmark. The main text states that Table 1 uses the Huang et al. (2022) PCACC ground truth, while Supplementary B.1 states that Table 1 uses the ProxyTTA ground truth. The reported nuScenes numbers for the proposed method (RMSE 1.561) match the PCACC column in Table S1 (RMSE 1.516) rather than the ProxyTTA column (RMSE 5.876), and on the ProxyTTA ground truth the proposed method is worse than the ProxyTTA baseline (5.876 vs. 5.509). The authors must state exactly which ground truth is used in each table, recompute the '21% average performance improvement' with a consistent benchmark choice, and explain the discrepancy.
  4. [Algorithm 1, line 8] The outlier-filtering algorithm as written is not reproducible. Line 8 assigns the set of deviating points by |y_hat_i - y_i| > tau, but Omega(y*) is defined as the set of reliable sparse point locations. To filter outliers, the condition should be |y_hat_i - y_i| <= tau, or the definition of Omega(y*) must be changed. This is a load-bearing component of the method and needs to be corrected in the manuscript.
  5. [Supplementary A.1 and Section 3.4] The method uses several dataset-dependent hyperparameters: lambda_smooth and lambda_rssim are 'adjusted according to the dataset', high-resolution images are optimized at 2x downsampling, the superpixel count is fixed at 200, and the optimization loop uses 200 iterations. For a method whose headline claim is zero-shot generalization, the paper should provide a single default configuration or show that the reported results are stable across a reasonable range of these hyperparameters. Otherwise, the per-dataset tuning itself becomes a contributor to the reported improvements, and the 'zero-shot' claim is weakened.
minor comments (6)
  1. [Eq. (9)] The notation in Eq. (9) is confusing: the text says the score is rewritten in terms of z0(zt), but the displayed update is for z_hat_t and the gradient is taken with respect to z_t. Please clarify which variable is optimized and how the two-step process in Eqs. (7)-(8) replaces this update.
  2. [Eq. (10)] Eq. (10) is not a well-defined mathematical statement as written; the left- and right-hand sides appear to involve scores at different arguments with no explicit distributions. Please specify the variables and the sense in which the approximation holds.
  3. [Algorithm 1, line 7] The input to the RANSAC regressor is written ambiguously as RANSAC Regressor(1_{Omega(y_i)} circle D_r, y_i). Please clarify whether the relative depth map is the input feature and the sparse metric depth is the target, and state how the filtering threshold tau is chosen.
  4. [Table 2] The inference-time comparison reports only seconds without specifying the GPU model and memory; please add the hardware configuration so the efficiency comparison is interpretable.
  5. [Section 3.4, Eq. (13)] The R-SSIM loss uses the depth estimated by an off-the-shelf model as d1; please state explicitly which model produces d1 and whether it is the same diffusion prior used for sampling, to rule out a concern that the regularizer is circularly pulling toward the prior's own output.
  6. [General] There are several typographical and phrasing issues, including 'alignend' in Section 3.2, 'embeded' in Related Work, 'naively' in the Conclusion, and an incomplete sentence in the Fig. 4 caption. A careful proofread is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are supported by external priors and test-time optimization against sparse measurements, not by fitting the reported benchmarks.

full rationale

I walked the claimed derivation chain. The method takes an external pretrained affine-invariant depth diffusion model (Marigold or DepthFM) as the score prior and, at test time, optimizes z0(zt) to satisfy sparse metric measurements (Eqs. 7-9). The reported gains are measured against external test-time adaptation baselines on held-out domain datasets, and no parameter is fitted to those benchmark numbers. The compatibility claim in Eq. 10 is supported by the reconstruction experiment in Supplementary Table S2, which empirically checks whether normalized metric depth maps can be re-represented by the prior via encoding, noise addition, and reverse sampling. That is a real, if limited, empirical test rather than a definitional identity; the main weakness, that the test degrades at high noise and does not fully cover random-noise trajectories, is an evidence gap rather than circularity. The R-SSIM loss explicitly regularizes the completed depth toward the structure of the prior output, which is a disclosed loss design, not a hidden prediction. The only apparent self-citation (Uni-DVPS for 3D scene understanding motivation) is not load-bearing. Overall, no step reduces to its own input by construction, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

All the load-bearing components are either standard diffusion theory or empirical assumptions; the only paper-specific constructs are the R-SSIM loss and the two-step alignment correction. No new physical entities are introduced. The free parameters are hand-tuned per dataset, which weakens the zero-shot claim.

free parameters (5)
  • Lambda_smooth (Eq. 14) = 0.2, adjusted per dataset
    Weight for local smoothness loss; hand-set per dataset.
  • Lambda_rssim (Eq. 14) = 0.3, adjusted per dataset
    Weight for R-SSIM structural loss; hand-set per dataset.
  • Inner optimization iterations (Eq. 7) = 200
    Number of gradient steps in the test-time alignment loop; fixed but not justified systematically.
  • Superpixel segment count N (Algorithm 1) = 200
    Number of segments for outlier filtering; hand-chosen.
  • Outlier threshold tau (Algorithm 1) = not reported numerically
    Deviation threshold for classifying outliers; not given in the paper.
assumptions (4)
  • standard math Score-based diffusion models provide a valid generative prior and the pretrained Marigold/DepthFM networks can act as the score function for depth.
    Assumed in Section 3.1; standard diffusion model theory.
  • domain assumption The normalized metric depth space lies within the support of the affine-invariant depth diffusion prior.
    Central premise of Section 3.2; supported only by the reconstruction experiment in Supplementary Table S2.
  • domain assumption Sparse depth measurements are synchronized with the RGB image and their noise can be handled by the prior-based outlier filtering.
    Assumed in Sections 3.2 and 3.3; real sensors may violate this.
  • ad hoc to paper Using the latent decoder D and remapping with noise in Eq. 8 keeps the optimized latent on the correct noise level manifold.
    The two-step hard alignment (Eq. 7-8) is an ad hoc correction; the paper does not prove that remapping yields a valid latent for the next reverse step.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior." pith.science (2026). https://pith.science/paper/XM2PTG76

@misc{pith2026250206338,
  author       = {Pith},
  title        = {Pith review of: Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XM2PTG76}},
  note         = {Machine review of arXiv:2502.06338}
}
read the original abstract

Depth completion, predicting dense depth maps from sparse depth measurements, is an ill-posed problem requiring prior knowledge. Recent methods adopt learning-based approaches to implicitly capture priors, but the priors primarily fit in-domain data and do not generalize well to out-of-domain scenarios. To address this, we propose a zero-shot depth completion method composed of an affine-invariant depth diffusion model and test-time alignment. We use pre-trained depth diffusion models as depth prior knowledge, which implicitly understand how to fill in depth for scenes. Our approach aligns the affine-invariant depth prior with metric-scale sparse measurements, enforcing them as hard constraints via an optimization loop at test-time. Our zero-shot depth completion method demonstrates generalization across various domain datasets, achieving up to a 21\% average performance improvement over the previous state-of-the-art methods while enhancing spatial understanding by sharpening scene details. We demonstrate that aligning a monocular affine-invariant depth prior with sparse metric measurements is a proven strategy to achieve domain-generalizable depth completion without relying on extensive training data. Project page: https://hyoseok1223.github.io/zero-shot-depth-completion/.

Figures

Figures reproduced from arXiv: 2502.06338 by the authors.

Figure 1
Figure 1. 3D-lifted depth completion results in out-of-domain cases. Regardless of supervised (Zhang et al. 2023) or unsupervised methods (Wong and Soatto 2021), most depth completion models perform poorly on out-of-domain data. In contrast, our zero-shot depth completion method, which employs test-time alignment, consistently achieves robust results. In this example, the other models are trained on the KITTI Depth Completion… view at source ↗
Figure 2
Figure 2. Illustration of our approach. At test time, we align the depth affinity from the prior (dashed lines) with the sparse depth measurements as a hard constraint (bold lines). This alignment propagates measurements across the scene to complete unobservable depth values. diffusion model that comprehends depth prior (Ke et al. 2024; Gui et al. 2024) by aligning it with sparse depth measure￾ment using the proposed test tim… view at source ↗
Figure 3
Figure 3. Test-time alignment process. We incorporate a two-step hard alignment process into the reverse sampling process including an optimization loop and resample at regular intervals. We optimize z0(zt) and remap it to zˆt. The latent is then decoded into depth, where the loss is measured against sparse depth. For visibility, the sparse depth points are enlarged. 3.2 Test-time Alignment with Hard Constraints Depth measure… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Alignment with metric depth. We evaluate our method’s effectiveness against ground truth (GT), accumulated semi-densely. We use only sparse depth (a) to align with actual metric depth values in complex scenes, ensuring a desirable solution. The white lines in (b), (c),…
Figure 5
Figure 5. Figure 5: Qualitative comparison on the nuScenes test set. In outdoor scenarios, our test-time alignment method performs robustly even under extreme weather conditions, clearly identifying critical elements such as vehicles and signs. RGB GT BN Adapt CoTTA ProxyTTA Ours [PITH_F…
Figure 6
Figure 6. Figure 6: Qualitative comparison on the NYU test set. In indoor scenarios, our test-time alignment method accurately captures scene structures (e.g., chairs) compared to the existing test-time adaptation methods. Method N-shot Scenario RMSE MAE VPP4DC 0 0.247 0.077 DepthPromptin…
Figure 8
Figure 8. Figure 8: Ablation studies. We ablate our proposed methods including test-time alignment, R-SSIM loss, and prior-based outlier filtering, to demonstrate their effectiveness. w/o R-SSIM Loss w/ R-SSIM Loss RGB [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Qualitative ablation of R-SSIM loss. This struc￾tural regularization sharpens details in areas such as signposts and car shapes. Acknowledgement We thank the members of AMILab (AMILab 2024) for their helpful discussions and proofreading. This work was sup￾ported by the…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 66 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Achanta, R.; Shaji, A.; Smith, K.; Lucchi, A.; Fua, P.; and Süsstrunk, S. 2012. SLIC Superpixels Compared to State-of-the-Art Superpixel Methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(11): 2274--2282

    work page 2012

  4. [4]

    AMILab. 2024. https://ami.postech.ac.kr/members https://ami.postech.ac.kr/members

    work page 2024

  5. [5]

    Bansal, A.; Chu, H.-M.; Schwarzschild, A.; Sengupta, S.; Goldblum, M.; Geiping, J.; and Goldstein, T. 2024. Universal Guidance for Diffusion Models. In Int. Conf. Learn. Represent

    work page 2024

  6. [6]

    Bartolomei, L.; Poggi, M.; Conti, A.; Tosi, F.; and Mattoccia, S. 2024. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In International Conference on 3D Vision (3DV), 1360--1370. IEEE

    work page 2024

  7. [7]

    H.; Vora, S.; Liong, V

    Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; ; and Beijbom, O. 2020. nuscenes: A multi- modal dataset for autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog

    work page 2020

  8. [8]

    Caron, M.; Touvron, H.; Misra, I.; Jégou, H.; Mairal, J.; Bojanowski, P.; and Joulin, A. 2021. Emerging Properties in Self-Supervised Vision Transformers. In Int. Conf. Comput. Vis

    work page 2021

Show all 73 references
  1. [9]

    Cheng, X.; Wang, P.; and Yang, R. 2018. Learning Depth with Convolutional Spatial Propagation Network. In Eur. Conf. Comput. Vis

    2018

  2. [10]

    Choe, J.; Im, S.; Rameau, F.; Kang, M.; and Kweon, I. S. 2021. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Int. Conf. Comput. Vis., 16086--16095

    2021

  3. [11]

    T.; Klasky, M

    Chung, H.; Kim, J.; Mccann, M. T.; Klasky, M. L.; and Ye, J. C. 2023. Diffusion Posterior Sampling for General Noisy Inverse Problems. In Int. Conf. Learn. Represent

    2023

  4. [12]

    Chung, H.; Lee, S.; and Ye, J. C. 2024. Decomposed Diffusion Sampler for Accelerating Large-Scale Inverse Problems. In Int. Conf. Learn. Represent

    2024

  5. [13]

    Chung, H.; Sim, B.; Ryu, D.; and Ye, J. C. 2022. Improving Diffusion Models for Inverse Problems using Manifold Constraints. In Adv. Neural Inform. Process. Syst

    2022

  6. [14]

    Chung, H.; Sim, B.; and Ye, J. C. 2022. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In IEEE Conf. Comput. Vis. Pattern Recog., 12413--12422

    2022

  7. [15]

    Conti, A.; Poggi, M.; Aleotti, F.; and Mattoccia, S. 2022. Unsupervised confidence for LiDAR depth maps and applications. In IEEE/RSJ International Conference on Intelligent Robots and Systems

    2022

  8. [16]

    Dhariwal, P.; and Nichol, A. 2021. Diffusion Models Beat GANs on Image Synthesis. In Adv. Neural Inform. Process. Syst

    2021

  9. [17]

    Efron, B. 2011. Tweedie’s Formula and Selection Bias. Journal of the American Statistical Association, 106(496): 1602--1614. PMID: 22505788

    2011

  10. [18]

    Eigen, D.; Puhrsch, C.; and Fergus, R. 2014. Depth map prediction from a single image using a multi-scale deep network. In Adv. Neural Inform. Process. Syst

    2014

  11. [19]

    A.; and Bolles, R

    Fischler, M. A.; and Bolles, R. C. 1981. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381–395

    1981

  12. [20]

    S.; Prestel, U.; Ma, P.; Kotovenko, D.; Grebenkova, O.; Baumann, S

    Gui, M.; Fischer, J. S.; Prestel, U.; Ma, P.; Kotovenko, D.; Grebenkova, O.; Baumann, S. A.; Hu, V. T.; and Ommer, B. 2024. DepthFM: Fast Monocular Depth Estimation with Flow Matching. arXiv:2403.13788

    2024 arXiv

  13. [21]

    Harris, C.; Stephens, M.; et al. 1988. A combined corner and edge detector. In Alvey vision conference, volume 15, 10--5244. Citeseer

    1988

  14. [22]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 770--778

    2016

  15. [23]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising Diffusion Probabilistic Models. In Adv. Neural Inform. Process. Syst

    2020

  16. [24]

    Ho, J.; and Salimans, T. 2022. Classifier-Free Diffusion Guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications

    2022

  17. [25]

    Hu, M.; Wang, S.; Li, B.; Ning, S.; Fan, L.; and Gong, X. 2021. PENet: Towards Precise and Efficient Image Guided Depth Completion. In IEEE International Conference on Robotics and Automation

    2021

  18. [26]

    Huang, S.; Gojcic, Z.; Huang, J.; and Andreas Wieser, K. S. 2022. Dynamic 3D Scene Analysis by Point Cloud Accumulation. In Eur. Conf. Comput. Vis

    2022

  19. [27]

    Ilg, E.; Cicek, O.; Galesso, S.; Klein, A.; Makansi, O.; Hutter, F.; and Brox, T. 2018. Uncertainty estimates and multi-hypotheses networks for optical flow. In Eur. Conf. Comput. Vis., 652--667

    2018

  20. [28]

    Ji-Yeon, K.; Hyun-Bin, O.; Byung-Ki, K.; Kim, D.; Kwon, Y.; and Oh, T.-H. 2024. Uni-DVPS: Unified Model for Depth-Aware Video Panoptic Segmentation. IEEE Robotics and Automation Letters, 9(7): 6186--6193

    2024

  21. [29]

    V.; Schindler, K.; and Obukhov, A

    Jia, Y.; Hoyer, L.; Huang, S.; Wang, T.; Gool, L. V.; Schindler, K.; and Obukhov, A. 2024. DGInStyle: Domain-Generalizable Semantic Segmentation with Image Diffusion Models and Stylized Semantic Control. In European Conference on Computer Vision, ECCV

    2024

  22. [30]

    Kam, J.; Kim, J.; Kim, S.; Park, J.; and Lee, S. 2022. CostDCNet: Cost Volume Based Depth Completion for a Single RGB-D Image. In Eur. Conf. Comput. Vis., 257--274. Springer

    2022

  23. [31]

    C.; and Schindler, K

    Ke, B.; Obukhov, A.; Huang, S.; Metzger, N.; Daudt, R. C.; and Schindler, K. 2024. Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation. In IEEE Conf. Comput. Vis. Pattern Recog

    2024

  24. [32]

    Y.; Chung, H.; and Ye, J

    Kim, J.; Park, G. Y.; Chung, H.; and Ye, J. C. 2024. Regularization by Texts for Latent Diffusion Inverse Solvers. arXiv:2311.15658

    2024 arXiv

  25. [33]

    Lee, H.-Y.; Tseng, H.-Y.; Lee, H.-Y.; and Yang, M.-H. 2024. Exploiting Diffusion Prior for Generalizable Dense Prediction. In IEEE Conf. Comput. Vis. Pattern Recog

    2024

  26. [34]

    Li, Z.; and Chen, J. 2015. Superpixel segmentation using Linear Spectral Clustering. In IEEE Conf. Comput. Vis. Pattern Recog., 1356--1363

    2015

  27. [35]

    Lin, Y.; Cheng, T.; Zhong, Q.; Zhou, W.; and Yang, H. 2022. Dynamic Spatial Propagation Network for Depth Completion. In AAAI

    2022

  28. [36]

    Lipson, L.; Teed, Z.; and Deng, J. 2021. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In International Conference on 3D Vision (3DV), 218--227. IEEE

    2021

  29. [37]

    D.; Gu, J.; Zhong, G.; Yang, M.-H.; and Kautz, J

    Liu, S.; Mello, S. D.; Gu, J.; Zhong, G.; Yang, M.-H.; and Kautz, J. 2017. Learning Affinity via Spatial Propagation Networks. In Adv. Neural Inform. Process. Syst

    2017

  30. [38]

    Liu, S.; Zeng, Z.; Ren, T.; Li, F.; Zhang, H.; Yang, J.; Li, C.; Yang, J.; Su, H.; Zhu, J.; and Zhang, L. 2023. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. arXiv:2303.05499

    2023 arXiv

  31. [39]

    Lopez-Rodriguez, A.; Busam, B.; and Mikolajczyk, K. 2020. Project to Adapt: Domain Adaptation for Depth Completion from Noisy and Sparse Sensor Data. In Asian Conf. Comput. Vis

    2020

  32. [40]

    Lu, K.; Barnes, N.; Anwar, S.; and Zheng, L. 2020. From depth what can you see? Depth completion via auxiliary image reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., 11306--11315

    2020

  33. [41]

    V.; and Karaman, S

    Ma, F.; Cavalheiro, G. V.; and Karaman, S. 2019. Self-supervised Sparse-to-Dense: Self-supervised Depth Completion from LiDAR and Monocular Camera. In IEEE International Conference on Robotics and Automation

    2019

  34. [42]

    Ma, F.; and Karaman, S. 2018. Sparse-to-Dense: Depth Prediction from Sparse Depth Samples and a Single Image. In IEEE International Conference on Robotics and Automation

    2018

  35. [43]

    McCormac, J.; Handa, A.; Leutenegger, S.; and Davison, A. J. 2017. Scenenet rgb-d: 5m photorealistic images of synthetic indoor trajectories with ground truth. In Int. Conf. Comput. Vis

    2017

  36. [44]

    Namekata, K.; Sabour, A.; Fidler, S.; and Kim, S. W. 2024. EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models. In Int. Conf. Learn. Represent

    2024

  37. [45]

    Park, H.; Gupta, A.; and Wong, A. 2024. Test-Time Adaptation for Depth Completion. In IEEE Conf. Comput. Vis. Pattern Recog

    2024

  38. [46]

    Park, J.; Joo, K.; Hu, Z.; Liu, C.-K.; and Kweon, I. S. 2020. Non-Local Spatial Propagation Network for Depth Completion. In Eur. Conf. Comput. Vis

    2020

  39. [47]

    Park, J.-H.; and Jeon, H.-G. 2024. A Simple yet Universal Framework for Depth Completion. In Adv. Neural Inform. Process. Syst

    2024

  40. [48]

    Park, J.-H.; Jeong, C.; Lee, J.; and Jeon, H.-G. 2024. Depth Prompting for Sensor-Agnostic Depth Estimation. In IEEE Conf. Comput. Vis. Pattern Recog

    2024

  41. [49]

    Qiu, J.; Cui, Z.; Zhang, Y.; Zhang, X.; Liu, S.; Zeng, B.; and Pollefeys, M. 2019. Deeplidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In IEEE Conf. Comput. Vis. Pattern Recog., 3313--3322

    2019

  42. [50]

    Ranftl, R.; Lasinger, K.; Hafner, D.; Schindler, K.; and Koltun, V. 2022. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. IEEE Trans. Pattern Anal. Mach. Intell., 44(3)

    2022

  43. [51]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. In IEEE Conf. Comput. Vis. Pattern Recog

    2022

  44. [52]

    Silberman, N.; Hoiem, D.; Kohli, P.; and Fergus, R. 2012. Indoor segmentation and support inference from rgbd images. In Eur. Conf. Comput. Vis., 746--760. Springer

    2012

  45. [53]

    M.; Zhang, Z.; Hu, X.; Qu, Q.; and Shen, L

    Song, B.; Kwon, S. M.; Zhang, Z.; Hu, X.; Qu, Q.; and Shen, L. 2024. Solving Inverse Problems with Latent Diffusion Models via Hard Data Consistency. In Int. Conf. Learn. Represent

    2024

  46. [54]

    Song, J.; Meng, C.; and Ermon, S. 2021. Denoising Diffusion Implicit Models. In Int. Conf. Learn. Represent

    2021

  47. [55]

    Song, Y.; Dhariwal, P.; Chen, M.; and Sutskever, I. 2023. Consistency models. In Int. Conf. Mach. Learn

    2023

  48. [56]

    P.; Kumar, A.; Ermon, S.; and Poole, B

    Song, Y.; Sohl-Dickstein, J.; Kingma, D. P.; Kumar, A.; Ermon, S.; and Poole, B. 2021. Score-Based Generative Modeling through Stochastic Differential Equations. In Int. Conf. Learn. Represent

    2021

  49. [57]

    Sun, P.; Kretzschmar, H.; Dotiwalla, X.; Chouard, A.; Patnaik, V.; Tsui, P.; Guo, J.; Zhou, Y.; Chai, Y.; Caine, B.; Vasudevan, V.; Han, W.; Ngiam, J.; Zhao, H.; Timofeev, A.; Ettinger, S.; Krivokon, M.; Gao, A.; Joshi, A.; Zhang, Y.; Shlens, J.; Chen, Z.; and Anguelov, D. 202...

    2020

  50. [58]

    Tang, J.; Tian, F.-P.; Feng, W.; Li, J.; and Tan, P. 2020. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30: 1116--1129

    2020

  51. [59]

    Uhrig, J.; Schneider, N.; Schneider, L.; Franke, U.; Brox, T.; and Geiger, A. 2017. Sparsity Invariant CNNs. In International Conference on 3D Vision (3DV)

    2017

  52. [60]

    Viereck, U.; Pas, A.; Saenko, K.; and Platt, R. 2017. Learning a visuomotor controller for real world robotic grasping using simulated depth images. In Conference on robot learning, 291--300. PMLR

    2017

  53. [61]

    Wang, D.; Shelhamer, E.; Liu, S.; Olshausen, B.; and Darrell, T. 2021. Tent: Fully Test-Time Adaptation by Entropy Minimization. In Int. Conf. Learn. Represent

    2021

  54. [62]

    Wang, J.; Yue, Z.; Zhou, S.; Chan, K. C. K.; and Loy, C. C. 2023 a . Exploiting Diffusion Prior for Real-World Image Super-Resolution. arXiv:2305.07015

    2023 arXiv

  55. [63]

    Wang, Q.; Fink, O.; Van Gool, L.; and Dai, D. 2022. Continual Test-Time Domain Adaptation. In IEEE Conf. Comput. Vis. Pattern Recog

    2022

  56. [64]

    Wang, Y.; Li, B.; Zhang, G.; Liu, Q.; Gao, T.; and Dai, Y. 2023 b . LRRU: Long-short Range Recurrent Updating Networks for Depth Completion. In Int. Conf. Comput. Vis

    2023

  57. [65]

    Wang, Z.; Bovik, A.; Sheikh, H.; and Simoncelli, E. 2004. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600--612

    2004

  58. [66]

    Wong, A.; Cicek, S.; and Soatto, S. 2021. Learning Topology From Synthetic Data for Unsupervised Depth Completion. IEEE Robotics and Automation Letters, 6(2): 1495--1502

    2021

  59. [67]

    Wong, A.; Fei, X.; Tsuei, S.; and Soatto, S. 2020. Unsupervised Depth Completion From Visual Inertial Odometry. IEEE Robotics and Automation Letters, 5(2): 1899--1906

    2020

  60. [68]

    Wong, A.; and Soatto, S. 2021. Unsupervised Depth Completion with Calibrated Backprojection Layers. In Int. Conf. Comput. Vis

    2021

  61. [69]

    Yang, L.; Kang, B.; Huang, Z.; Xu, X.; Feng, J.; and Zhao, H. 2024. Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data. In IEEE Conf. Comput. Vis. Pattern Recog

    2024

  62. [70]

    Yang, L.; Zhang, Z.; Song, Y.; Hong, S.; Xu, R.; Zhao, Y.; Zhang, W.; of Methods, C. M. A. C. S.; Applications, B.; and Yang, M.-H. 2023. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1--39

    2023

  63. [71]

    Yin, W.; Zhang, C.; Chen, H.; Cai, Z.; Yu, G.; Wang, K.; Chen, X.; and Shen, C. 2023. Metric3D: Towards Zero-shot Metric 3D Prediction from A Single Image. In Int. Conf. Comput. Vis

    2023

  64. [72]

    Zhang, Y.; Guo, X.; Poggi, M.; Zhu, Z.; Huang, G.; and Mattoccia, S. 2023. Completionformer: Depth completion with convolutions and vision transformers. In IEEE Conf. Comput. Vis. Pattern Recog

    2023

  65. [73]

    Zhao, L.; Zheng, W.; Duan, Y.; Zhou, J.; and Lu, J. 2024. SPTR: Structure-Preserving Transformer for Unsupervised Indoor Depth Completion. IEEE Transactions on Circuits and Systems for Video Technology, 34(4): 2439--2452

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.