Pith. sign in

REVIEW 4 major objections 5 minor 73 references

Perfecting Depth: Uncertainty-Aware Enhancement of Metric Depth

T0 review · 4 major / 5 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 This paper establishes that sample variance from a diffusion model trained on clean depth and run on raw sensor depth marks unreliable pixels, enabling a deterministic second stage to output dense, accurate metric depth.

desk verdict Clever two-stage idea with a real mechanism gap: variance-to-reliability is asserted, not proven, and the DIODE evaluation is biased—still worth refereeing. read the letter →

arxiv 2506.04612 v1 pith:MRABKGP3 submitted 2025-06-05 cs.CV

Jinyoung Jun , Lei Chu , Jiahao Li , Yan Lu , Chang-Su Kim This is my paper
classification cs.CV
keywords sensordepthenhancementdiffusionmodelsuncertaintyestimationtraining-inferencegapcompletioninpaintingmetricmaskedspatialpropagation
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 proposes a two-stage pipeline, called Perfecting Depth, that turns a diffusion model's sampling uncertainty into a per-pixel reliability map for raw sensor depth. In the first stage, a diffusion model trained only on clean synthetic depth is run repeatedly on the raw, imperfect depth map; the pixel-wise variance across these reconstructions is read as an estimate of how unreliable each measurement is. In the second stage, a deterministic masked-spatial-propagation network refines only the high-variance regions, restoring metric scale by least-squares alignment to the trusted pixels. The paper claims that no hand-crafted artifact prior is needed, and that a model trained solely on synthetic data generalizes to real sensor noise, missing regions, and downstream depth tasks. If correct, this gives autonomous driving, robotics, and augmented-reality systems a way to clean imperfect depth hardware without retraining per sensor.

What carries the argument

The load-bearing mechanism is the training-inference domain gap inside the diffusion model, paired with multi-sample posterior variance. Concretely, the diffusion denoiser is trained on clean depth with the conditioning input formed by masking the clean depth; during inference the conditioning input is the raw sensor depth. Repeating the denoising ten times from distinct noise initializations yields samples whose per-pixel mean and variance are treated as geometric cue and uncertainty map, respectively. A fixed threshold converts the variance into a certainty mask; least-squares fitting rescales the diffusion mean to the reliable pixels, and a masked spatial propagation network iteratively refines the uncertain regions while leaving trusted measurements intact.

What would settle it

Take a clean depth map with known ground truth, corrupt a fixed subset of pixels with realistic sensor artifacts, run the stochastic stage with its suggested ten samples, and rank pixels by variance. If the variance ranking does not concentrate on the corrupted pixels—say, its area under the ROC curve for detecting them is near chance—then the variance-reliability link fails and the refinement stage would be driven by noise rather than by a trustworthy uncertainty signal.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a deliberately induced training-inference mismatch converts a diffusion model into an uncertainty sensor for depth. The model learns the conditional distribution of clean depth on clean data with random masks; at test time the conditioning depth is set to the raw sensor depth, so the model sees a conditioning signal unlike anything in training. Across ten stochastic denoising runs, the sample variance at each pixel is claimed to be sharply peaked where the sensor value agrees with the learned posterior and broad where it does not, so high variance marks missing or artifact-corrupted pixels. Those pixels are then handed to a deterministic refinement network, which propagates the trusted values using the RGB image, the filtered reliable depth, the diffusion mean (after least-squares scale and shift recovery), and the variance map itself as guidance. The result is claimed to be a dense, artifact-free depth map that preserves edge detail, recovers metric scale without a fixed training range, and transfers from synthetic indoor training to real-world indoor benchmarks.

Load-bearing premise

The entire uncertainty mechanism rests on the assumption that a diffusion model trained only on clean depth will, when fed a noisy or missing sensor value, produce a wider spread of reconstructions at exactly the pixels where the sensor is wrong.

Editorial extensions

If this is right

  • Feeding the refined depth to downstream relative-depth estimators improves their fine-tuning RMSE and rank correlation on both noisy and clean real-world benchmarks.
  • The pipeline consistently improves noisy depth completion outputs at 5, 10, and 20 percent noise ratios, with the reported RMSE improvement over the masked-propagation baseline growing as noise increases.
  • The stochastic stage alone acts as a depth inpainter: across hole sizes up to half the image, it outperforms monocular depth estimators, with its largest advantage on small holes.
  • Because the refinement network is trained with random scale and shift augmentation and recovers metric scale from trusted pixels, the framework transfers across datasets with different depth ranges without re-training.
  • Training both stages exclusively on synthetic data with no artifact labels is sufficient to detect and repair real-world sensor artifacts, so the approach needs no per-sensor noise model.

Reading between the lines

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

  • A direct test of the mechanism would compare this diffusion variance map against an ensemble or MC-dropout uncertainty map on the same corrupted depth inputs; if the training-inference gap is what carries the signal, the diffusion variance should separate valid from invalid pixels substantially better than an in-distribution ensemble.
  • The fixed variance threshold that works across very different sensors in the paper suggests a learned, per-scene adaptive threshold as a natural extension, one the paper does not explore.
  • Because the generative prior is learned from indoor synthetic scenes, out-of-distribution content such as sky, mirrors, and transparent objects is the likely failure mode; the paper itself notes an infinite-depth normalization limitation.
  • The same clean-trained, raw-conditioned recipe could be applied to other sensor modalities, such as LiDAR intensity or time-of-flight confidence maps, wherever a clean-data generative prior exists; this is an extrapolation, not a paper claim.
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. The paper proposes a two-stage framework, Perfecting Depth, for sensor depth enhancement. The first stage trains a diffusion model on clean masked synthetic depth and, at inference, conditions on raw noisy depth; N stochastic reconstructions are drawn and the per-pixel sample variance sigma-hat^2 is used as an uncertainty map. The second stage uses a deterministic masked spatial propagation network (MSPN) to refine high-variance regions while propagating reliable depth values, with a least-squares fit to restore metric scale. Experiments are reported on DIODE-Indoor, NYUv2, and ScanNet for sensor depth enhancement, noisy depth completion, and depth inpainting, with the claim that synthetic-only training generalizes to real-world data.

Significance. If the central mechanism is sound, the framework is a significant step toward general sensor-depth enhancement: it avoids hand-crafted artifact priors, combines stochastic uncertainty estimation with deterministic refinement, and demonstrates synthetic-only training with cross-dataset transfer. The paper is clearly written and the experimental suite is broad, including downstream fine-tuning of relative depth estimators and controlled noisy-completion benchmarks. The main weakness is that the load-bearing variance-to-reliability link is not established: the theoretical derivation in Appendix A is internally inconsistent, and the key DIODE-Indoor evaluation filters test labels with the method's own uncertainty map, which cannot validate that high variance marks sensor artifacts. The paper also does not quantify the contribution of sigma-hat^2 to the refinement stage. These issues are addressable with additional experiments, but they currently prevent the central claim from being fully supported.

major comments (4)
  1. [Appendix A, Eqs. (15)-(18)] The Bayes derivation is internally inconsistent. Eq. (16) assumes I and M are independent of D_true to factorize the likelihood, yet Eq. (18) reintroduces p(D_true | I, M) as the prior. Since the independence assumption was used to eliminate I and M from the likelihood, the final expression contradicts that assumption. Therefore the theoretical support for the variance-reliability monotonicity is invalid. Please either provide a correct derivation or explicitly present Eq. (8) as a heuristic and support the monotonicity claim with direct empirical validation.
  2. [Section 4.3, Table 2] The DIODE-Indoor evaluation filters test labels using sigma-hat^2 from the first stage of the proposed pipeline. This makes the evaluation circular for validating the uncertainty map: the mask used for evaluation is produced by the method being evaluated. The improvement over 'Raw data' on this filtered set does not establish that high-variance regions correspond to sensor artifacts. Please report results on the full DIODE-Indoor validation set without filtering, or use known artifact masks (e.g., from synthetic corruptions) to measure the detection accuracy of sigma-hat^2 against ground-truth invalid pixels.
  3. [Section 3.1, Eq. (8) and Remark] The claimed monotonicity that sample variance is low for reliable pixels and high for unreliable pixels is not established. Because the model is trained with Dcond = masked clean depth, it may learn to treat provided non-zero depth values as trustworthy and reproduce them with low variance even when they are artifacts at inference. The paper's own Remark states that filtering 'primarily' focuses on missing areas, not non-zero artifacts. The paper needs direct evidence that sigma-hat^2 distinguishes non-zero artifacts from valid depth, for example by injecting known non-zero corruptions and reporting detection rates, or by showing refinement accuracy as a function of sigma-hat^2 ranking.
  4. [Section 4.6, Fig. 6(c)] The ablation for the use of sigma-hat^2 in the refinement stage is only qualitative. Quantify the contribution of sigma-hat^2 to final depth accuracy, for example by comparing RMSE and delta-1.25 with and without sigma-hat^2 in the guidance input on the noisy depth completion benchmark of Section 4.4. Additionally, report the sensitivity of the full pipeline to the threshold epsilon on a quantitative metric, rather than only showing masked-region visualizations.
minor comments (5)
  1. [Section 4.4] The text 'improves MSPN's RMSE performance by 289%' is numerically incorrect: the RMSE decreases from 0.655 to 0.227, which is a 65% relative reduction, not a 289% improvement. Please correct this phrasing.
  2. [Captions, Figures 5 and 10] The word 'correponding' should be 'corresponding' in both figure captions.
  3. [Section 4.5] 'exisiting' is a typo for 'existing'.
  4. [Section 3.2] The sentence 'To train our deterministic refinement network, we also use the outputs of our diffusion model also on synthetic data' is grammatically awkward and should be rephrased.
  5. [Notation throughout] The notation for ground-truth depth is inconsistent (Dtrue vs. D_true). Please unify the notation.

Circularity Check

2 steps flagged · score 5.0 of 10

Appendix A's variance-reliability 'derivation' is circular: Eq. 16 assumes I,M independent of Dtrue, then Eq. 18 re-inserts p(Dtrue|I,M) by definition; the DIODE evaluation also filters test labels with the same σ̂² it validates. The empirical pipeline has independent external benchmarks, so the circularity is partial.

  1. self definitional [Appendix A, Eqs. 16-18 (and Eq. 8, Sec. 3.1 'Variance vs. Reliability')]
    "p(Dcond, I, M | Dtrue,(i,j)) ≈ p(Dcond | Dtrue,(i,j)) · p(I, M) ... At this stage, we recognize that p(Dtrue,(i,j)), the prior on the ground-truth depth, can itself depend on I and M ... Therefore, we refine the prior term as: p(Dtrue,(i,j)) = p(Dtrue,(i,j) | I, M)."

    Eq. 16 assumes I and M are independent of Dtrue so that p(I,M) factors out and cancels, leaving Eq. 17 with a bare prior p(Dtrue). The very next step redefines that prior as p(Dtrue | I, M), re-introducing by definition the dependence just assumed away. Eq. 18 is therefore not derived from Bayes' rule and the stated assumptions; it is the target posterior decomposition inserted by hand. The subsequent claim that reliable pixels yield sharply peaked likelihoods and unreliable pixels broad posteriors is an unproved assertion about the learned model, so the theoretical support for the variance-reliability link assumes the link rather than establishing it.

  2. other [Sec. 4.3, Sensor depth enhancement, evaluation protocol]
    "Since the test labels in DIODE-Indoor are noisy as well, we filter out unreliable pixels using σ̂2 from the first stage of our pipeline and measure metric depth performance."

    The DIODE metric is computed only on pixels that the method's own σ̂² labels as reliable. The same σ̂² defines the certainty mask M_σ̂² (Eq. 9) that drives the refinement, so the model output under test also selects the test subset. Without a known artifact mask, this protocol cannot validate the central claim that σ̂² detects invalid pixels, and the reported DIODE comparison is partly self-selecting. The paper's own Remark concedes that σ̂²-based filtering 'primarily focuses on the missing areas,' further weakening the artifact-detection claim.

full rationale

Most of the pipeline is not circular: σ̂² is measured from N stochastic samples of the diffusion model (Eq. 5), the scale and shift a,b are per-image least-squares fits to Drel (Eq. 11), and the refinement network is supervised on Hypersim ground truth. The reuse of MSPN [30] is component adoption from the authors' prior work with external validation, not load-bearing self-citation. However, the paper's theoretical grounding of the central variance-to-reliability mechanism is circular (Appendix A: independence assumption in Eq. 16, then p(Dtrue):=p(Dtrue|I,M) in Eq. 18), and the DIODE-Indoor evaluation filters its test labels with the same σ̂² that the paper is validating. Because the method is still tested against external ground truth in NYUv2/ScanNet inpainting and in the controlled noisy-depth-completion setting, the central empirical contribution retains independent content. The circularity is therefore partial rather than total.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central mechanism rests on the unverified assumption that diffusion-sample variance is a reliable artifact indicator. The paper's 'theoretical analysis' is a heuristic Bayes argument with a contradictory independence assumption. Two important hyperparameters (epsilon, N) are hand-selected, and no code is released.

free parameters (2)
  • epsilon (uncertainty threshold) = 0.01
    Eq. 9 defines the certainty mask threshold; Sec. 4.6 sets epsilon = 0.01; Fig. 6(b) justifies it empirically. This threshold controls which pixels are refined and is set by hand.
  • N (number of diffusion samples) = 10
    Sec. 4.6 states that beyond 10 samples the gains are minimal but inference time increases. This is a hand-chosen trade-off.
assumptions (4)
  • ad hoc to paper Posterior factorization: p(D_cond | D_true, I, M) approx p(D_cond | D_true), and I and M are independent of D_true
    Invoked in Appendix A Eqs. 16-18 to derive the variance-reliability link; the paper then re-introduces an I,M-dependent prior, making the assumption inconsistent.
  • domain assumption The diffusion model's posterior variance is high for sensor-invalid pixels and low for valid pixels
    The central mechanism, asserted in Sec. 3.1 'Variance vs. Reliability', is not guaranteed by the training objective; it is empirically assumed.
  • domain assumption Synthetic Hypersim distribution is close enough to real indoor scenes for the learned prior to be informative
    Sec. 3.1 Remarks and Sec. 4.1: the model is trained only on Hypersim but applied to DIODE, NYUv2, and ScanNet; the domain shift is assumed small.
  • domain assumption Sensor noise N_sensor is localized and most pixels are accurate (Eq. 1)
    The method assumes sparse localized corruption; if noise is global or correlated, the uncertainty map and refinement strategy may fail.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Perfecting Depth: Uncertainty-Aware Enhancement of Metric Depth." pith.science (2026). https://pith.science/paper/MRABKGP3

@misc{pith2026250604612,
  author       = {Pith},
  title        = {Pith review of: Perfecting Depth: Uncertainty-Aware Enhancement of Metric Depth},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRABKGP3}},
  note         = {Machine review of arXiv:2506.04612}
}
read the original abstract

We propose a novel two-stage framework for sensor depth enhancement, called Perfecting Depth. This framework leverages the stochastic nature of diffusion models to automatically detect unreliable depth regions while preserving geometric cues. In the first stage (stochastic estimation), the method identifies unreliable measurements and infers geometric structure by leveraging a training-inference domain gap. In the second stage (deterministic refinement), it enforces structural consistency and pixel-level accuracy using the uncertainty map derived from the first stage. By combining stochastic uncertainty modeling with deterministic refinement, our method yields dense, artifact-free depth maps with improved reliability. Experimental results demonstrate its effectiveness across diverse real-world scenarios. Furthermore, theoretical analysis, various experiments, and qualitative visualizations validate its robustness and scalability. Our framework sets a new baseline for sensor depth enhancement, with potential applications in autonomous driving, robotics, and immersive technologies.

Figures

Figures reproduced from arXiv: 2506.04612 by the authors.

Figure 1
Figure 1. Sensor depth enhancement result on DIODE-Indoor. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed framework. inherent data noise. Early works used Bayesian approxima￾tions [23, 32, 34] and ensembles [2–4, 10, 11, 14, 36, 48] to capture epistemic uncertainty. Our approach, similar to con￾current work [54], generates an ensemble through multiple denoising iterations of a single diffusion model, rather than by training multiple models. Different with these methods, the epistemic uncertai… view at source ↗
Figure 4
Figure 4. Refinement result on noisy depth completion using [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison of depth inpainting results on ScanNet. For each depth map, the correponding error map is provided [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: (a) Visualization of σˆ 2 and the corresponding masked depth map according to N. (b) Masking result according to thresh￾old ϵ. (c) Efficacy of σˆ 2 for refinement. such cases, one potential alternative is to leverage a monoc￾ular relative depth estimator, where known d…
Figure 7
Figure 7. Figure 7: Sensor depth enhancement result on DIODE-Indoor. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Refinement result on noisy depth completion using CFormer. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Refinement result on noisy depth completion using MSPN. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparison of depth inpainting results on ScanNet. For each depth map, the correponding error map is provided [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison of depth inpainting results on NYUv2. For each depth map, the correponding error map is provided [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

73 extracted references · 57 canonical work pages

  1. [54]

    Zero-shot uncertainty quantification using diffusion probabilistic models

    Dule Shu and Amir Barati Farimani. Zero-shot uncertainty quantification using diffusion probabilistic models. arXiv preprint arXiv:2408.04718, 2024. 3

    arXiv 2024

  2. [1]

    GAUDI: A neural architect for immersive 3d scene genera- tion

    Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Wal- ter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. GAUDI: A neural architect for immersive 3d scene genera- tion. In NeurIPS, 2022. 2

    work page 2022

  3. [2]

    Escaping the sample trap: Fast and accurate epistemic uncertainty estimation with pairwise-distance estimators

    Lucas Berry and David Meger. Escaping the sample trap: Fast and accurate epistemic uncertainty estimation with pairwise-distance estimators. 2023. 3

    work page 2023

  4. [3]

    Normalizing flow ensem- bles for rich aleatoric and epistemic uncertainty modeling

    Lucas Berry and David Meger. Normalizing flow ensem- bles for rich aleatoric and epistemic uncertainty modeling. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 6806–6814, 2023

    work page 2023

  5. [4]

    Shedding light on large generative networks: Estimating epistemic uncer- tainty in diffusion models

    Lucas Berry, Axel Brando, and David Meger. Shedding light on large generative networks: Estimating epistemic uncer- tainty in diffusion models. In The 40th Conference on Un- certainty in Artificial Intelligence, 2024. 3

    work page 2024

  6. [5]

    AdaBins: Depth estimation using adaptive bins

    Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. AdaBins: Depth estimation using adaptive bins. In CVPR,

    work page

  7. [6]

    Align your latents: High-resolution video synthesis with la- tent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with la- tent diffusion models. In CVPR, 2023. 2

    2023

  8. [7]

    Vir- tual kitti 2

    Yohann Cabon, Naila Murray, and Martin Humenberger. Vir- tual kitti 2. arXiv preprint arXiv:2001.10773, 2020. 5

    arXiv 2001

Show all 73 references
  1. [8]

    Depth esti- mation via affinity learned with convolutional spatial propa- gation network

    Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth esti- mation via affinity learned with convolutional spatial propa- gation network. In ECCV, 2018. 2

    2018

  2. [9]

    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. 2

    2019

  3. [10]

    Waic, but why? generative ensembles for robust anomaly detection

    Hyunsun Choi, Eric Jang, and Alexander A Alemi. Waic, but why? generative ensembles for robust anomaly detection. arXiv preprint arXiv:1810.01392, 2018. 3

    2018 arXiv

  4. [11]

    Deep reinforcement learning in a handful of trials using probabilistic dynamics models

    Kurtland Chua, Roberto Calandra, Rowan McAllister, and Sergey Levine. Deep reinforcement learning in a handful of trials using probabilistic dynamics models. Advances in neural information processing systems, 31, 2018. 3

    2018

  5. [12]

    Spar- sity agnostic depth completion

    Andrea Conti, Matteo Poggi, and Stefano Mattoccia. Spar- sity agnostic depth completion. In WACV, 2023. 2

    2023

  6. [13]

    Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 5, 12, 14

    2017

  7. [14]

    Decomposition of un- certainty in bayesian deep learning for efficient and risk- sensitive learning

    Stefan Depeweg, Jose-Miguel Hernandez-Lobato, Finale Doshi-Velez, and Steffen Udluft. Decomposition of un- certainty in bayesian deep learning for efficient and risk- sensitive learning. In International conference on machine learning, pages 1184–1193. PMLR, 2018. 3

    2018

  8. [15]

    BERT: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional trans- formers for language understanding. In NAACL, 2018. 2

    2018

  9. [16]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2

    2021

  10. [17]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  11. [18]

    Depthlab: Real-time 3d in- teraction with depth maps for mobile augmented reality

    Ruofei Du, Eric Turner, Maksym Dzitsiuk, Luca Prasso, Ivo Duarte, Jason Dourgarian, Joao Afonso, Jose Pascoal, Josh Gladstone, Nuno Cruces, et al. Depthlab: Real-time 3d in- teraction with depth maps for mobile augmented reality. In UIST, 2020. 1

    2020

  12. [19]

    Depth map prediction from a single image using a multi-scale deep net- work

    David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep net- work. In NeurIPS, 2014. 1

    2014

  13. [20]

    Structure and content-guided video synthesis with diffusion models

    Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, 2023. 2

    2023

  14. [21]

    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 CVPR, 2018. 1

    2018

  15. [22]

    Geowiz- ard: Unleashing the diffusion priors for 3d geometry estima- tion from a single image

    Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowiz- ard: Unleashing the diffusion priors for 3d geometry estima- tion from a single image. In ECCV, 2024. 1, 2

    2024

  16. [23]

    Deep bayesian active learning with image data

    Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International conference on machine learning, pages 1183–1192. PMLR,

  17. [24]

    Are we ready for autonomous driving? the KITTI vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, 2012. 1

    2012

  18. [25]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2

    2022

  19. [26]

    Towards fast and accurate real-world depth super- resolution: Benchmark dataset and baseline

    Lingzhi He, Hongguang Zhu, Feng Li, Huihui Bai, Runmin Cong, Chunjie Zhang, Chunyu Lin, Meiqin Liu, and Yao Zhao. Towards fast and accurate real-world depth super- resolution: Benchmark dataset and baseline. In CVPR, 2021. 2

    2021

  20. [27]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. In NeurIPS, 2020. 2

    2020

  21. [28]

    Hamid Izadinia, Qi Shan, and Steven M. Seitz. IM2CAD. In CVPR, 2017. 1

    2017

  22. [29]

    DDP: Diffusion model for dense visual prediction

    Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. DDP: Diffusion model for dense visual prediction. In ICCV, 2023. 2

    2023

  23. [30]

    Masked spatial propagation network for sparsity-adaptive depth re- finement

    Jinyoung Jun, Jae-Han Lee, and Chang-Su Kim. Masked spatial propagation network for sparsity-adaptive depth re- finement. In CVPR, 2024. 2, 4, 5, 6, 7, 12, 14

    2024

  24. [31]

    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 CVPR, 2024. 1, 2, 4, 7, 8, 12

    2024

  25. [32]

    What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3 9

    2017

  26. [33]

    Maurice G. Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81–93, 1938. 6

    1938

  27. [34]

    Batchbald: Efficient and diverse batch acquisition for deep bayesian active learning

    Andreas Kirsch, Joost Van Amersfoort, and Yarin Gal. Batchbald: Efficient and diverse batch acquisition for deep bayesian active learning. Advances in neural information processing systems, 32, 2019. 3

    2019

  28. [35]

    Continuously masked transformer for image inpainting

    Keunsoo Ko and Chang-Su Kim. Continuously masked transformer for image inpainting. In ICCV, 2023. 8

    2023

  29. [36]

    Simple and scalable predictive uncertainty estima- tion using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estima- tion using deep ensembles. Advances in neural information processing systems, 30, 2017. 3

    2017

  30. [37]

    Single-image depth estimation based on Fourier domain analysis

    Jae-Han Lee, Minhyeok Heo, Kyung-Rae Kim, and Chang- Su Kim. Single-image depth estimation based on Fourier domain analysis. In CVPR, 2018. 1

    2018

  31. [38]

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

    Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 1

    1907 arXiv

  32. [39]

    Mat: Mask-aware transformer for large hole image in- painting

    Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image in- painting. In CVPR, 2022. 8

    2022

  33. [40]

    Depthlab: From partial to complete

    Zhiheng Liu, Ka Leong Cheng, Qiuyu Wang, Shuzhe Wang, Hao Ouyang, Bin Tan, Kai Zhu, Yujun Shen, Qifeng Chen, and Ping Luo. Depthlab: From partial to complete. arXiv preprint arXiv:2412.18153, 2024. 2, 4

    2024

  34. [41]

    Sparse-to-dense: Depth prediction from sparse depth samples and a single image

    Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, 2018. 2

    2018

  35. [42]

    Guided depth super-resolution by deep anisotropic diffusion

    Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Guided depth super-resolution by deep anisotropic diffusion. In CVPR, 2023. 2

    2023

  36. [43]

    Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition

    Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In CVPR, 2024. 2

    2024

  37. [44]

    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. 2

    2022

  38. [45]

    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 ECCV, 2020. 2

    2020

  39. [46]

    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 CVPR, 2024. 2

    2024

  40. [47]

    Ecodepth: Effective conditioning of diffusion models for monocular depth estimation

    Suraj Patni, Aradhye Agarwal, and Chetan Arora. Ecodepth: Effective conditioning of diffusion models for monocular depth estimation. In CVPR, 2024. 2

    2024

  41. [48]

    The hidden uncertainty in a neural networks ac- tivations

    Janis Postels, Hermann Blum, Yannick Str ¨umpler, Cesar Cadena, Roland Siegwart, Luc Van Gool, and Federico Tombari. The hidden uncertainty in a neural networks ac- tivations. arXiv preprint arXiv:2012.03082, 2020. 3

    2012 arXiv

  42. [49]

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

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocu- lar depth estimation: Mixing datasets for zero-shot cross- dataset transfer. IEEE Trans. Pattern Anal. Mach. Intell., 44 (3):1623–1637, 2020. 1

    2020

  43. [50]

    Hypersim: A photorealistic syn- thetic dataset for holistic indoor scene understanding

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic syn- thetic dataset for holistic indoor scene understanding. In ICCV, 2021. 5, 12

    2021

  44. [51]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 2, 5

    2022

  45. [52]

    The surprising effectiveness of diffusion models for optical flow and monocular depth estimation

    Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. In NeurIPS, 2024. 2

    2024

  46. [53]

    Laion-5B: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 5

    2022

  47. [55]

    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 ECCV, 2012. 1, 5, 12, 14

    2012

  48. [56]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In ICLR, 2021. 2

    2021

  49. [57]

    Dif- fusion models for monocular depth estimation: Overcoming challenging conditions

    Fabio Tosi, Pierluigi Zama Ramirez, and Matteo Poggi. Dif- fusion models for monocular depth estimation: Overcoming challenging conditions. In ECCV, 2024. 2

    2024

  50. [58]

    Maxvit: Multi-axis vision transformer

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 5, 12

    2022

  51. [59]

    Dai, Andrea F

    Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Moham- madreza Mostajabi, Steven Basart, Matthew R. Walter, and Gregory Shakhnarovich. DIODE: A dense indoor and out- door depth dataset. arXiv preprint arXiv:1908.00463, 2...

    1908 arXiv

  52. [60]

    High-fidelity pluralistic image completion with transform- ers

    Ziyu Wan, Jingbo Zhang, Dongdong Chen, and Jing Liao. High-fidelity pluralistic image completion with transform- ers. In ICCV, 2021. 8

    2021

  53. [61]

    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 ICCV, 2023. 2

    2023

  54. [62]

    Deep3D: Fully automatic 2D-to-3D video conversion with deep convolu- tional neural networks

    Junyuan Xie, Ross Girshick, and Ali Farhadi. Deep3D: Fully automatic 2D-to-3D video conversion with deep convolu- tional neural networks. In ECCV, 2016. 1

    2016

  55. [63]

    SimMIM: A simple framework for masked image modeling

    Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A simple framework for masked image modeling. In CVPR, 2022. 2

    2022

  56. [64]

    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 CVPR, 2024. 1, 5, 6, 7, 8

    2024

  57. [65]

    Depth any- thing v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2. arXiv:2406.09414, 2024. 5, 6, 7, 8 10

    2024 arXiv

  58. [66]

    Aggregating feature point cloud for depth completion

    Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In ICCV, 2023. 2

    2023

  59. [67]

    3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models

    Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Trans. Graph., 42(4):1–16, 2023. 2

    2023

  60. [68]

    Joint- net: Extending text-to-image diffusion for dense distribution modeling

    Jingyang Zhang, Shiwei Li, Yuanxun Lu, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, and Yao Yao. Joint- net: Extending text-to-image diffusion for dense distribution modeling. In ICLR, 2023. 2

    2023

  61. [69]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2

    2023

  62. [70]

    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 CVPR, 2023. 2, 6, 7, 14

    2023

  63. [71]

    Unleashing text-to-image diffusion models for visual perception

    Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In ICCV, 2023. 2

    2023

  64. [72]

    Discrete cosine transform network for guided depth map super-resolution

    Zixiang Zhao, Jiangshe Zhang, Shuang Xu, Zudi Lin, and Hanspeter Pfister. Discrete cosine transform network for guided depth map super-resolution. In CVPR, 2022. 2

    2022

  65. [73]

    Spherical space feature decomposition for guided depth map super-resolution

    Zixiang Zhao, Jiangshe Zhang, Xiang Gu, Chengli Tan, Shuang Xu, Yulun Zhang, Radu Timofte, and Luc Van Gool. Spherical space feature decomposition for guided depth map super-resolution. In CVPR, 2023. 2 11 A. Posterior approximation In this section, we will discuss about why t...

    2023

Pith tools

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