Pith. sign in

REVIEW 3 major objections 6 minor 56 references

Self-Supervised Enhancement for Depth from a Lightweight ToF Sensor with Monocular Images

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Self-supervised depth enhancement lifts 8×8 ToF depth to near-supervised accuracy.

desk verdict A solid self-supervised framework for 8x8 ToF enhancement with good ablations, but the results all rest on a simulated sensor; read the numbers as a proof of concept until real L5 data appears. read the letter →

arxiv 2506.13444 v2 pith:DCM4KAPH submitted 2025-06-16 cs.CV

Laiyan Ding , Hualie Jiang , Jiwei Chen , Rui Huang This is my paper · ORCID
classification cs.CV
keywords self-superviseddepthestimationenhancementTime-of-Flightsensorlow-resolutionsparsecompletionsubmanifoldconvolutionguidedfeaturefusionscalerecovery
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

SelfToF claims that an off-the-shelf, 8×8 Time-of-Flight sensor can produce detailed, metric-accurate depth maps when its readings are fused with ordinary monocular RGB video, and that no ground-truth depth is needed for training. The paper starts from a self-supervised depth-estimation pipeline, feeds the ToF depth into both the depth and pose networks, adds a zone-wise depth consistency loss, and recovers metric scale with a median-of-median scaling step. On the standard indoor benchmark, the enhancement cuts absolute relative error from 0.154 (image-only baseline) to 0.049, a result close to a supervised counterpart's 0.046. A sparsity-robust variant, SelfToF*, holds near that accuracy even when up to 40% of ToF zones are missing. The reason to care: lightweight ToF sensors cost a few dollars and run at a fraction of a watt, so removing the label requirement makes high-quality depth accessible wherever RGB video already exists.

What carries the argument

The load-bearing machinery is a two-stream DepthNet (RGB encoder plus a small 8×8 depth encoder) and a two-stream PoseNet, trained jointly by view synthesis. Three losses drive learning: a photometric loss warping the source frame into the target, an edge-aware smoothness loss on disparity, and a zone-wise depth consistency loss that compares Gaussian mean and variance per zone between the sensor and the prediction. Metric scale is recovered by median-of-median scaling (MMS): compute, for each of the 64 zones, the ratio of the sensor's zone median to the predicted zone median, then take the median of those ratios as the global scale. For the sparsity-robust SelfToF*, the depth encoder is rebuilt with submanifold sparse convolutions so invalid zones never propagate features, and a guided feature fusion module flattens RGB and depth features at 8×8, generates an affinity map from position-encoded RGB features, propagates depth features along it, and sums the original RGB, original depth, and propagated depth features. The affinity mechanism is the piece that lets each invalid zone borrow statistics from the most similar valid zones.

What would settle it

Run SelfToF on real data from the low-cost 8×8 ToF sensor, with the camera co-registered to RGB and trained on the paper's simulated data. If the real-sensor absolute relative error is more than roughly twice the simulated-test error on matched indoor scenes, the Gaussian-per-zone, aligned, no-cutoff simulation is not faithful, and the reported gains would not transfer to hardware.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that an 8×8 depth map carrying only per-zone mean and variance contains enough metric information to disambiguate the scale of a self-supervised depth network, while the network's photometric supervision supplies the detail. Adding the ToF readings as a second input stream to DepthNet, penalizing the difference between predicted and sensed zone distributions with a Gaussian consistency loss, recovering scale by the median of per-zone median ratios, and feeding the ToF depth to PoseNet together lower abs rel on the indoor benchmark from 0.154 to 0.049, and the test-time output is scale-aware without any further scaling. The paper also claims that when some zones are invalid, a submanifold sparse depth encoder plus an RGB-guided, position-aware feature fusion keeps accuracy nearly intact: at 40% missing zones, SelfToF* reports 0.061 abs rel versus 0.070 for SelfToF. Crucially, the best self-supervised model trails the supervised same-architecture baseline by only a small margin (0.049 versus 0.046), suggesting the ground-truth dependency is nearly eliminated.

Load-bearing premise

The training and evaluation assume that the simulated ToF signal—a Gaussian fitted per 8×8 zone, perfectly aligned with the RGB image, with no maximum-distance cutoff—faithfully represents what the real low-cost sensor returns, including its noise, multipath, and missing-zone behavior.

Editorial extensions

If this is right

  • Any device that already has an RGB camera and a few-dollar ToF sensor can act as a metric depth camera without collecting depth labels.
  • The self-supervised training signal is unlabeled video, so the method can be retrained on a new deployment site's own footage rather than requiring a capture rig.
  • Because ToF depth also goes into the PoseNet, ego-motion estimation improves as a by-product (rotational error on the generalization dataset drops from 1.740° to 1.521°).
  • Partial sensor failure is survivable: as long as roughly 60% of zones remain valid, the submanifold and guided-fusion variant keeps quality nearly unchanged.

Reading between the lines

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

  • The median-of-median scaling trick is not specific to ToF; any sparse depth sensor that can report per-cell medians (lidar rings, stereo disparity bins) could reuse MMS to make a self-supervised network metric-aware.
  • A hardware test on the actual low-cost sensor is the natural next experiment; the paper's own simulation assumption is the main risk to the generality of the reported numbers.
  • The guided feature fusion is effectively a tiny attention module at 8×8 resolution; pairing it with larger RGB backbones or depth foundation models is a plausible route to sharper boundaries and better generalization.
  • If real sensor noise breaks the Gaussian assumption, the depth consistency loss could be weakened or replaced with a distribution-free zone comparison, and MMS would still carry the scale recovery.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This paper addresses depth enhancement for low-cost lightweight ToF sensors (8x8 zones) using paired RGB images in a self-supervised manner. The proposed framework, SelfToF, extends a monocular self-supervised depth estimation baseline (Monodepth2) by adding the low-resolution ToF depth as an additional input to both the DepthNet and PoseNet, a depth-consistency loss that compares per-zone Gaussian statistics, and a median-of-median scaling module for metric scale recovery. A variant SelfToF* uses a submanifold depth encoder and guided feature fusion to handle missing zones. Training uses only monocular sequences and the sensor's low-resolution depth, without dense ground truth. On simulated NYU data, SelfToF reduces absolute relative error from 0.154 (baseline) to 0.049; it also shows favorable generalization to ScanNet and robustness to zone sparsity. Code is publicly available.

Significance. If the results transfer to real hardware, the method could make lightweight ToF-based depth sensing with metric scale accessible in self-supervised settings, which is practically important for low-cost IoT and robotics. The architecture is technically sound, with an ablation chain (Table I) that traces the contribution of each design choice; the release of code is commendable; and the method demonstrates robustness to a simulated sparsity setting. The main reason to temper the significance is that all quantitative evidence is produced on simulated ToF signals with explicit assumptions (zone-RGB alignment, no maximum-distance cutoff) that are not validated on a real L5 sensor, so the magnitude of the reported gains may not hold in practice.

major comments (3)
  1. [Sec. IV.A and Sec. III.B] All quantitative results (Tables I-VI) are obtained with simulated ToF signals produced by fitting Gaussians per 8x8 zone following DELTAR, not with data from a real L5 sensor. Section III.B explicitly assumes 'that L5 zones are aligned with RGB images and no maximum distance constraint exists,' and the conclusions concede that real-world datasets are future work. Because the two most load-bearing components—the depth-consistency loss (Eq. 6) and the scale-recovery module (Eq. 2)—depend critically on the reliability of per-zone ToF statistics, the central claim that SelfToF enhances depth from a lightweight ToF sensor is not directly supported by the presented evidence. Please either add real-sensor validation (or a carefully designed hardware-in-the-loop simulation) or scope the claims to simulated settings and analyze the risks of misalignment and missing zones.
  2. [Sec. IV.E, Table II] The performance comparison does not include DELTAR, the prior method specifically designed for the same L5 depth-enhancement task. The strongest learning-based baseline in the table is MonoDepth2, and the only supervised reference is the authors' own architecture with ground-truth supervision (last row of Table I). Since DELTAR is the established method for this task, its absence makes it difficult to assess the gap between the proposed self-supervised approach and the state of the art, and weakens the claim that the method is 'effective' relative to previous architectures. Please add a comparison with DELTAR (or a similar supervised L5 method) or justify its omission.
  3. [Sec. IV.G] The sparsity experiments randomly zero out zones, which is an artificial pattern. Real L5 sensors typically exhibit invalid zones that are spatially clustered and range-dependent (e.g., due to specular surfaces or out-of-range measurements). The robustness of SelfToF* is therefore demonstrated only against a specific synthetic sparsity model. Evaluate with structured sparsity (e.g., contiguous blocks) or real-data artifacts, and report performance accordingly.
minor comments (6)
  1. [Tables I-VI] Tables I-VI report a single training run; no error bars or repeated runs are given, making it hard to judge the significance of small differences (e.g., Table V, SelfToF+SDE vs SelfToF+GFF). Please add variance estimates at least for the main results.
  2. [Eq. (6)] Eq. (6) is described as 'the distance between two 1D Gaussian distributions' with reference to Wasserstein metrics; however, the expression is the sum of squared differences of means and standard deviations, not the (squared) 2-Wasserstein distance as usually defined. Please clarify or adjust the citation.
  3. [Sec. III.B.1] The MAC count comparison (2.38G vs 2.59G) should specify the input resolution and the exact configuration to allow reproduction.
  4. [Sec. IV.F, Table III] The evaluation protocol for scale-ambiguous baselines in the ScanNet generalization (Table III) is not described; please state whether median scaling is applied to the baseline in that table, as it is in Table I.
  5. [Figure 7] In Figure 7, the color coding of valid/invalid zone boundaries and the visualization of the affinity map should be explained in the caption to make the qualitative claim more checkable.
  6. [Abstract and Eq. (2)] The abstract contains a grammar error: 'SelfToF* maintain robust performance' should be 'SelfToF* maintains robust performance.' Additionally, please clarify whether the median operation in Eq. (2) is applied over all zones including invalid (zero-valued) zones, and how the scale recovery behaves when more than 50% of zones are invalid (which would give a zero median).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SelfToF uses the ToF input as a sensor signal and as a consistency target, not as a disguised groundtruth, so the reported derivation is self-contained.

full rationale

The derivation chain is self-contained. SelfToF's training objective (Eq. 7) combines photometric warping (Eqs. 3-4), edge-aware smoothness (Eq. 5), and a zone-wise Gaussian depth-consistency loss (Eq. 6) that compares the predicted depth's per-zone mean and variance with the input ToF signal. This is a standard sensor-consistency loss, not a hidden groundtruth; with weight wdc = 0.01, the network is not forced to reproduce the input. At test time, MMS (Eq. 2) multiplies the scale-ambiguous network output by median(C_i^t / median(d_i^t)), which is a legitimate use of the sensor input to recover metric scale rather than a fitted parameter disguised as a prediction; the dense high-resolution structure still comes from the RGB image and the network, so the output is not the input by construction. Comparisons against MonoDepth2, nearest-neighbor upsampling, and guided filtering are external baselines, and the ToF simulation follows DELTAR [4], an external benchmark. The only self-citations ([13], [19], [22]) appear for data-sampling details, prior baselines, or related works and are not load-bearing for the central claim. The paper does state a clear limitation in Sec. III.B: 'Assuming that L5 zones are aligned with RGB images and no maximum distance constraint exists,' and Sec. V calls for 'collecting large-scale real-world datasets' in future work. These are external-validity caveats about the sim-to-real gap, not evidence that any claimed prediction reduces to its inputs. No circular step is exhibited, so the appropriate finding is a score of 0.

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

The method rests on a sensor model taken from DELTAR and on the standard photometric-consistency assumption. The three loss weights and alpha are tuned hyperparameters. No new physical entities are introduced.

free parameters (2)
  • Loss weights (w_ph, w_s, w_dc) = 1.0, 0.1, 0.01
    Set empirically (Sec. III.C.4); the choice affects the balance between photometric, smoothness, and depth-consistency terms.
  • Photometric weight alpha = 0.85
    Conventional value from prior self-supervised depth works (Sec. III.C.1), not tuned for this task.
assumptions (4)
  • domain assumption L5 zones are aligned with RGB images and there is no maximum distance constraint
    Stated in Sec. III.B; real L5 has a limited field-of-view and max range, and alignment may require calibration, which the method does not address.
  • domain assumption Each ToF zone can be approximated by a fitted Gaussian distribution with mean and variance
    Borrowed from DELTAR [4]; used both to simulate training data (Sec. IV.A) and to define the depth consistency loss (Sec. III.C.3).
  • domain assumption Photometric consistency between temporally adjacent frames (Lambertian surfaces, static scene)
    Inherited from Monodepth2 [10]; the self-supervised signal breaks for moving objects, reflections, and occlusions.
  • domain assumption The simulated ToF data generated by Gaussian fitting on NYU/ScanNet depth is representative of real lightweight ToF output
    The whole evaluation rests on this; the authors note in Sec. V that real-world data is future work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Enhancement for Depth from a Lightweight ToF Sensor with Monocular Images." pith.science (2026). https://pith.science/paper/DCM4KAPH

@misc{pith2026250613444,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Enhancement for Depth from a Lightweight ToF Sensor with Monocular Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DCM4KAPH}},
  note         = {Machine review of arXiv:2506.13444}
}
read the original abstract

Depth map enhancement using paired high-resolution RGB images offers a cost-effective solution for improving low-resolution depth data from lightweight ToF sensors. Nevertheless, naively adopting a depth estimation pipeline to fuse the two modalities requires groundtruth depth maps for supervision. To address this, we propose a self-supervised learning framework, SelfToF, which generates detailed and scale-aware depth maps. Starting from an image-based self-supervised depth estimation pipeline, we add low-resolution depth as inputs, design a new depth consistency loss, propose a scale-recovery module, and finally obtain a large performance boost. Furthermore, since the ToF signal sparsity varies in real-world applications, we upgrade SelfToF to SelfToF* with submanifold convolution and guided feature fusion. Consequently, SelfToF* maintain robust performance across varying sparsity levels in ToF data. Overall, our proposed method is both efficient and effective, as verified by extensive experiments on the NYU and ScanNet datasets. The code is available at \href{https://github.com/denyingmxd/selftof}{https://github.com/denyingmxd/selftof}.

Figures

Figures reproduced from arXiv: 2506.13444 by the authors.

Figure 1
Figure 1. L5 sensing principle and example inputs and outputs with our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Training pipeline of our model. During forward, the model takes [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of multi-modality DepthNet and PoseNet. The fusion [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overview of guided feature fusion. We first downsample (DS) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Visual results and error maps on the NYU dataset. Using ToF as [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Visual results and error maps on NYU dataset. Compared with [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 36 canonical work pages

  1. [1]

    Sscnav: Confidence-aware semantic scene completion for visual semantic navigation,

    Y . Liang, B. Chen, and S. Song, “Sscnav: Confidence-aware semantic scene completion for visual semantic navigation,” in 2021 IEEE international conference on robotics and automation (ICRA) . IEEE, 2021, pp. 13 194–13 200

    work page 2021

  2. [2]

    Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai,

    T. Wang, X. Mao, C. Zhu, R. Xu, R. Lyu, P. Li, X. Chen, W. Zhang, K. Chen, T. Xue, et al. , “Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 19 757–19 767

    work page 2024

  3. [3]

    Stmicroelectronics: Time-of-flight (tof) 8x8 multizone ranging sensor with wide field of view

    “Stmicroelectronics: Time-of-flight (tof) 8x8 multizone ranging sensor with wide field of view.” [Online]. Available: https: //www.st.com/en/imaging-and-photonics-solutions/vl53l5cx.html

    work page

  4. [4]

    Deltar: Depth estimation from a light-weight tof sensor and rgb image,

    Y . Li, X. Liu, W. Dong, H. Zhou, H. Bao, G. Zhang, Y . Zhang, and Z. Cui, “Deltar: Depth estimation from a light-weight tof sensor and rgb image,” in European conference on computer vision . Springer, 2022, pp. 619–636

    work page 2022

  5. [5]

    Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor,

    X. Liu, Y . Li, Y . Teng, H. Bao, G. Zhang, Y . Zhang, and Z. Cui, “Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor,” in Proceedings of the ieee/cvf international conference on computer vision , 2023, pp. 1–11

    work page 2023

  6. [6]

    Sgnet: Structure guided network via gradient-frequency awareness for depth map super-resolution,

    Z. Wang, Z. Yan, and J. Yang, “Sgnet: Structure guided network via gradient-frequency awareness for depth map super-resolution,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 6, 2024, pp. 5823–5831

    2024

  7. [7]

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

    L. He, H. Zhu, F. Li, H. Bai, R. Cong, C. Zhang, C. Lin, M. Liu, and Y . Zhao, “Towards fast and accurate real-world depth super-resolution: Benchmark dataset and baseline,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 9229–9238

    work page 2021

  8. [8]

    Non-local spatial propagation network for depth completion,

    J. Park, K. Joo, Z. Hu, C.-K. Liu, and I. So Kweon, “Non-local spatial propagation network for depth completion,” in Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIII 16 . Springer, 2020, pp. 120–136

    work page 2020

Show all 56 references
  1. [9]

    Completionformer: Depth completion with convolutions and vision transformers,

    Y . Zhang, X. Guo, M. Poggi, Z. Zhu, G. Huang, and S. Mattoccia, “Completionformer: Depth completion with convolutions and vision transformers,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2023, pp. 18 527–18 536

    2023

  2. [10]

    Digging into self-supervised monocular depth estimation,

    C. Godard, O. Mac Aodha, M. Firman, and G. J. Brostow, “Digging into self-supervised monocular depth estimation,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 3828–3838

    2019

  3. [11]

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

    F. Ma, G. V . Cavalheiro, and S. Karaman, “Self-supervised sparse- to-dense: Self-supervised depth completion from lidar and monocular camera,” in 2019 International Conference on Robotics and Automa- tion (ICRA). IEEE, 2019, pp. 3288–3295

    2019

  4. [12]

    Depth completion towards different sensor configurations via relative depth map estimation and scale re- covery,

    Y . Long, H. Yu, and B. Liu, “Depth completion towards different sensor configurations via relative depth map estimation and scale re- covery,” Journal of Visual Communication and Image Representation , vol. 80, p. 103272, 2021

    2021

  5. [13]

    Plnet: Plane and line priors for unsupervised indoor depth estimation,

    H. Jiang, L. Ding, J. Hu, and R. Huang, “Plnet: Plane and line priors for unsupervised indoor depth estimation,” in 2021 International Conference on 3D Vision (3DV) . IEEE, 2021, pp. 741–750

    2021

  6. [14]

    Guided image filtering,

    K. He, J. Sun, and X. Tang, “Guided image filtering,” IEEE transac- tions on pattern analysis and machine intelligence , vol. 35, no. 6, pp. 1397–1409, 2012

    2012

  7. [15]

    Deep ordinal regression network for monocular depth estimation,

    H. Fu, M. Gong, C. Wang, K. Batmanghelich, and D. Tao, “Deep ordinal regression network for monocular depth estimation,” in Pro- ceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 2002–2011

    2018

  8. [16]

    Unsupervised cnn for single view depth estimation: Geometry to the rescue,

    R. Garg, V . K. Bg, G. Carneiro, and I. Reid, “Unsupervised cnn for single view depth estimation: Geometry to the rescue,” in Com- puter Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14 . Springer, 2016, pp. 740–756

    2016

  9. [17]

    Unsupervised learning of monocular depth estimation with bundle adjustment, super- resolution and clip loss,

    L. Zhou, J. Ye, M. Abello, S. Wang, and M. Kaess, “Unsupervised learning of monocular depth estimation with bundle adjustment, super- resolution and clip loss,” arXiv preprint arXiv:1812.03368 , 2018

    2018 arXiv

  10. [18]

    Unsupervised scale-consistent depth and ego-motion learning from monocular video,

    J. Bian, Z. Li, N. Wang, H. Zhan, C. Shen, M.-M. Cheng, and I. Reid, “Unsupervised scale-consistent depth and ego-motion learning from monocular video,” Advances in neural information processing systems, vol. 32, 2019

    2019

  11. [19]

    Unsupervised monocular depth perception: Focusing on moving objects,

    H. Jiang, L. Ding, Z. Sun, and R. Huang, “Unsupervised monocular depth perception: Focusing on moving objects,”IEEE Sensors Journal, vol. 21, no. 24, pp. 27 225–27 237, 2021

    2021

  12. [20]

    Depth prediction without the sensors: Leveraging structure for unsupervised learning from monocular videos,

    V . Casser, S. Pirk, R. Mahjourian, and A. Angelova, “Depth prediction without the sensors: Leveraging structure for unsupervised learning from monocular videos,” in Proceedings of the AAAI conference on artificial intelligence, vol. 33, no. 01, 2019, pp. 8001–8008

    2019

  13. [21]

    Self-supervised surround-view depth estimation with volumetric feature fusion,

    J.-H. Kim, J. Hur, T. P. Nguyen, and S.-G. Jeong, “Self-supervised surround-view depth estimation with volumetric feature fusion,” Ad- vances in Neural Information Processing Systems , vol. 35, pp. 4032– 4045, 2022

    2022

  14. [22]

    Towards cross-view- consistent self-supervised surround depth estimation,

    L. Ding, H. Jiang, J. Li, Y . Chen, and R. Huang, “Towards cross-view- consistent self-supervised surround depth estimation,” arXiv preprint arXiv:2407.04041, 2024

    2024 arXiv

  15. [23]

    Surrounddepth: Entangling surrounding views for self- supervised multi-camera depth estimation,

    Y . Wei, L. Zhao, W. Zheng, Z. Zhu, Y . Rao, G. Huang, J. Lu, and J. Zhou, “Surrounddepth: Entangling surrounding views for self- supervised multi-camera depth estimation,” in Conference on robot learning. PMLR, 2023, pp. 539–549

    2023

  16. [24]

    Exploring efficiency of vision transform- ers for self-supervised monocular depth estimation,

    A. Karpov and I. Makarov, “Exploring efficiency of vision transform- ers for self-supervised monocular depth estimation,” in 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR) . IEEE, 2022, pp. 711–719

    2022

  17. [25]

    Single image depth prediction with wavelet decom- position,

    M. Ramamonjisoa, M. Firman, J. Watson, V . Lepetit, and D. Tur- mukhambetov, “Single image depth prediction with wavelet decom- position,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 11 089–11 098

    2021

  18. [26]

    Moving indoor: Un- supervised video depth learning in challenging environments,

    J. Zhou, Y . Wang, K. Qin, and W. Zeng, “Moving indoor: Un- supervised video depth learning in challenging environments,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 8618–8627

    2019

  19. [27]

    P 2 net: Patch-match and plane- regularization for unsupervised indoor depth estimation,

    Z. Yu, L. Jin, and S. Gao, “P 2 net: Patch-match and plane- regularization for unsupervised indoor depth estimation,” in European Conference on Computer Vision . Springer, 2020, pp. 206–222

    2020

  20. [28]

    Auto-rectify network for unsupervised indoor depth estimation,

    J.-W. Bian, H. Zhan, N. Wang, T.-J. Chin, C. Shen, and I. Reid, “Auto-rectify network for unsupervised indoor depth estimation,”IEEE transactions on pattern analysis and machine intelligence , vol. 44, no. 12, pp. 9802–9813, 2021

    2021

  21. [29]

    Gasmono: Geometry-aided self-supervised monocular depth estima- tion for indoor scenes,

    C. Zhao, M. Poggi, F. Tosi, L. Zhou, Q. Sun, Y . Tang, and S. Mattoccia, “Gasmono: Geometry-aided self-supervised monocular depth estima- tion for indoor scenes,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 16 209–16 220

    2023

  22. [30]

    Unsupervised depth com- pletion from visual inertial odometry,

    A. Wong, X. Fei, S. Tsuei, and S. Soatto, “Unsupervised depth com- pletion from visual inertial odometry,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 1899–1906, 2020

    1906

  23. [31]

    Unsupervised depth completion with cali- brated backprojection layers,

    A. Wong and S. Soatto, “Unsupervised depth completion with cali- brated backprojection layers,” in Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision , 2021, pp. 12 747–12 756

    2021

  24. [32]

    Selfdeco: Self-supervised monocular depth completion in challenging indoor environments,

    J. Choi, D. Jung, Y . Lee, D. Kim, D. Manocha, and D. Lee, “Selfdeco: Self-supervised monocular depth completion in challenging indoor environments,” in 2021 ieee international conference on robotics and automation (icra). IEEE, 2021, pp. 467–474

    2021

  25. [33]

    Sparsity invariant cnns,

    J. Uhrig, N. Schneider, L. Schneider, U. Franke, T. Brox, and A. Geiger, “Sparsity invariant cnns,” in 2017 international conference on 3D Vision (3DV) . IEEE, 2017, pp. 11–20

    2017

  26. [34]

    Pixel-adaptive convolutional neural networks,

    H. Su, V . Jampani, D. Sun, O. Gallo, E. Learned-Miller, and J. Kautz, “Pixel-adaptive convolutional neural networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 11 166–11 175

    2019

  27. [35]

    Monitored distillation for positive congruent depth completion,

    T. Y . Liu, P. Agrawal, A. Chen, B.-W. Hong, and A. Wong, “Monitored distillation for positive congruent depth completion,” in European Conference on Computer Vision . Springer, 2022, pp. 35–53

    2022

  28. [36]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” Communications of the ACM , vol. 65, no. 1, pp. 99–106, 2021

    2021

  29. [37]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  30. [38]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255

    2009

  31. [39]

    Spatial transformer networks,

    M. Jaderberg, K. Simonyan, A. Zisserman, et al., “Spatial transformer networks,” Advances in neural information processing systems, vol. 28, 2015

    2015

  32. [40]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE transactions on image processing , vol. 13, no. 4, pp. 600–612, 2004

    2004

  33. [41]

    Learning depth from monocular videos using direct methods,

    C. Wang, J. M. Buenaposada, R. Zhu, and S. Lucey, “Learning depth from monocular videos using direct methods,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 2022–2030

    2018

  34. [42]

    A class of wasserstein metrics for probability distributions

    C. R. Givens and R. M. Shortt, “A class of wasserstein metrics for probability distributions.” Michigan Mathematical Journal , vol. 31, no. 2, pp. 231–240, 1984

    1984

  35. [43]

    Submanifold sparse convolutional networks,

    B. Graham and L. van der Maaten, “Submanifold sparse convolutional networks,” arXiv preprint arXiv:1706.01307 , 2017

    2017 arXiv

  36. [44]

    Spconv: Spatially sparse convolution library,

    S. Contributors, “Spconv: Spatially sparse convolution library,” https: //github.com/traveller59/spconv, 2022

    2022

  37. [45]

    Clustering by passing messages between data points,

    B. J. Frey and D. Dueck, “Clustering by passing messages between data points,” science, vol. 315, no. 5814, pp. 972–976, 2007

    2007

  38. [46]

    Adaptive affinity fields for semantic segmentation,

    T.-W. Ke, J.-J. Hwang, Z. Liu, and S. X. Yu, “Adaptive affinity fields for semantic segmentation,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 587–602

    2018

  39. [47]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Informa- tion Processing Systems , 2017

    2017

  40. [48]

    Indoor seg- mentation and support inference from rgbd images,

    N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor seg- mentation and support inference from rgbd images,” in Computer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12. Springer, 2012, pp. 746–760

    2012

  41. [49]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

    2019

  42. [50]

    Adam: A method for stochastic optimization,

    D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  43. [51]

    Adabins: Depth estimation using adaptive bins,

    S. F. Bhat, I. Alhashim, and P. Wonka, “Adabins: Depth estimation using adaptive bins,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 4009–4018

    2021

  44. [52]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 5828–5839

    2017

  45. [53]

    Monoindoor: Towards good practice of self-supervised monocular depth estimation for indoor environ- ments,

    P. Ji, R. Li, B. Bhanu, and Y . Xu, “Monoindoor: Towards good practice of self-supervised monocular depth estimation for indoor environ- ments,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12 787–12 796

    2021

  46. [54]

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

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 371–10 381

    2024

  47. [55]

    Repurposing diffusion-based image generators for monocular depth estimation,

    B. Ke, A. Obukhov, S. Huang, N. Metzger, R. C. Daudt, and K. Schindler, “Repurposing diffusion-based image generators for monocular depth estimation,” in Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) , 2024

    2024

  48. [56]

    Prompting depth anything for 4k resolution accurate metric depth estimation,

    H. Lin, S. Peng, J. Chen, S. Peng, J. Sun, M. Liu, H. Bao, J. Feng, X. Zhou, and B. Kang, “Prompting depth anything for 4k resolution accurate metric depth estimation,” arXiv preprint arXiv:2412.14015 , 2024

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.