Pith. sign in

REVIEW 2 major objections 7 minor 50 references

Region-aware Depth Scale Adaptation with Sparse Measurements

T0 review · 2 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A region-aware fitting procedure—one separate scale per segmented region, calibrated by sparse depth points—converts relative depth predictions into metric depth without retraining, using fewer measurements than global scaling baselines.

desk verdict Region-aware test-time scaling is a genuine and practically useful twist on global scale-shift, but the headline results rest entirely on ideal sparse-depth sampling. read the letter →

arxiv 2507.14879 v1 pith:3OAX4EFE submitted 2025-07-20 cs.CV

Rizhao Fan , Tianfang Ma , Zhigen Li , Ning An , Jian Cheng This is my paper
classification cs.CV
keywords monoculardepthestimationmetricrecoveryscaleadaptationfoundationmodelssparsemeasurementsregion-awarescalingsurfacefitting
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

Large single-image depth models give depth only up to an unknown scale, which blocks direct use in robotics, driving, and augmented reality. This paper claims that the missing metric scale can be recovered at test time from a small number of sparse depth measurements by treating the scene not as one global depth field but as a set of segmented regions, each with its own fitted scale and shift. Two fitting variants are tested: per-region linear scaling and per-region surface fitting that also absorbs local plane tilt. On the primary indoor benchmark and a second benchmark from a different sensor, the region-aware fits are reported to be substantially more accurate than global Median, Linear Fit, and low-cost-sensor scaling, reaching a mean absolute relative error of 0.031 and 0.050, respectively, for the strongest depth foundation model tested, even with only 250 to 2000 random depth points. If these results hold under real sensor noise, the method offers a training-free way to give relative-depth models a metric scale at deployment time.

What carries the argument

The carrying mechanism is a per-region least-squares fit driven by sparse metric depth points. Segmentation masks split the image into regions; inside each region the sparse measurements solve a small linear system—$D_m=\alpha D+\beta$ for the linear fit, or $z_1=\alpha z_2+\beta x+\gamma y+\delta$ for the surface fit—and the resulting parameters are applied to every pixel of that region's relative depth. When a region has too few measurement points, the algorithm repeatedly expands it to include neighboring regions until the fit is determined. This mechanism converts the global-scaling question into many small local calibration problems, which is what lets a few hundred points match or beat methods that consume the entire depth map.

What would settle it

Take the same two benchmarks and replace random ground-truth sampling with a realistic sensor mask—for example, a simulated 16-beam LiDAR pattern with depth-dependent noise and gaps between beams—then rerun SLF and SSF at 250 to 2000 points. If the reported mean absolute relative error values (0.039 and 0.031 for SSF-1000 and SSF-2000 on the primary benchmark) move toward the global Linear Fit baseline, the region-aware advantage depends on ideal sampling; if they stay near the reported numbers, the method survives real sensor constraints.

Watch

Extended reading notes

Core claim

The paper's central claim is that a predicted relative depth map behaves as a mosaic of local surfaces, so metric depth is recovered correctly only when each surface is rescaled independently rather than by one global scale-shift. To demonstrate this, the authors segment the image into regions, fit each region's sparse metric measurements against the relative prediction, and apply the fitted transform densely inside that region. The linear variant fits $D_m=\alpha D+\beta$ per region; the surface variant fits $z_1=\alpha z_2+\beta x+\gamma y+\delta$, reflecting the observation that a depth patch behaves locally like a tilted plane. Regions that lack enough sparse points are merged with neighboring regions until the fit is computable. The paper reports that this region-aware procedure outperforms global scaling methods on both benchmarks at every sampled measurement budget, that surface fitting beats linear fitting at equal point counts, and that accuracy improves steadily as the number of measurements grows from 250 to 2000.

Load-bearing premise

The sparse depth measurements are assumed to be synchronized with the image, dense enough in every region (directly or after expansion), and free of noise, because every fitted scale factor is computed from those points alone.

Editorial extensions

If this is right

  • Any depth foundation model can be given metric depth on a new scene from a cheap depth sensor's points, with no training data from that scene and no risk of forgetting the model's pretrained generalization.
  • With 250 random depth points, the region-aware fits already beat global transformations that use the full ground-truth depth map on the tested benchmarks.
  • The surface-fitting variant is consistently more accurate than the linear variant at the same sample count, so fitting local plane structure is better than fitting only scale and shift.
  • Accuracy rises monotonically from 250 to 2000 samples, so users can trade sensor density against accuracy at deployment time.
  • Applying the region-aware formulation to existing global baselines improves their accuracy, indicating that the benefit comes from the region decomposition itself.

Reading between the lines

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

  • Beyond the paper, the advantage of region-aware fitting should scale with scene heterogeneity: cluttered multi-object scenes should show the largest gains, while scenes close to a single plane should show little or none; that predicted domain-dependence is testable without new methodology.
  • Beyond the paper, realistic sparse sensors with beam patterns that miss whole objects will force the algorithm's region-expansion step to merge across object boundaries, so the fitted parameters for unmeasured objects become extrapolations from neighbors; measuring per-region error under beam-mask sampling rather than random sampling would quantify this.
  • Beyond the paper, the per-region fit can be turned into an active-sensing procedure: after an initial fit, the residual between predicted and fitted depth identifies the regions with the worst local parameters, and the next sensor measurements can be directed there to refine the scale map.
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

2 major / 7 minor

Summary. The paper proposes a training-free method to convert relative depth predictions from monocular depth estimation foundation models into metric scale using sparse depth measurements. The method segments the image with a foundation segmentation model, then fits each segmented region independently: SLF estimates a per-region scale and shift by linear regression, while SSF estimates a per-region surface model in (relative depth, x, y) by least squares, expanding a region through neighbors when too few sparse points are available. Experiments on NYUv2 and VOID with MiDaS, Depth Anything v1, and Depth Anything v2 report that SLF and SSF consistently outperform global scaling baselines, with SSF-2000 reaching Abs Rel 0.031 on NYUv2 and 0.050 on VOID for Depth Anything v2. The paper claims the approach preserves the generalization of the base models without retraining or fine-tuning.

Significance. If the results hold, the contribution is a simple, reproducible, and training-free alternative to test-time scale adaptation that improves over global fitting while leaving the base network untouched. The consistent ordering from global scaling to SLF to SSF across three backbone models and two datasets is a genuine strength, and the inclusion of ground-truth-based global baselines with a matched information source makes the core comparison meaningful. The method has no learned parameters and is straightforward to implement. Its significance is tempered by the idealized sparse-input assumption: all experiments use noiseless, uniformly sampled ground-truth points, while the motivating application is real LiDAR/ToF/stereo data with beam patterns, range-dependent noise, and missing regions. The absence of variance estimates over random sampling further limits the strength of the quantitative claims.

major comments (2)
  1. [Section 4.1, Tables 1 and 2] The central comparison is established only under idealized sparse measurements. Section 4.1 states that the 250, 500, 1000, and 2000 sparse points are randomly sampled from the ground truth data, meaning the inputs are noiseless, uniformly distributed, and never missing at object boundaries. Algorithm 1 depends on per-region least-squares fits and region expansion, and under realistic sensor patterns a region with no returns inherits a neighboring region's transform, which can be systematically wrong when the neighbor is a different surface. The paper already includes an LF-LiDAR baseline computed from a simulated beam pattern, but SLF and SSF are never evaluated on those or any other non-uniform or noisy patterns. I therefore do not consider the practical claim of bridging the gap to real deployment to be established. I would like to see experiments applying SLF and SSF to the existing LF-LiDAR beam simulator or to sparse maps with added Gaussian/impulse noise and missing regions, together with a report of how the gain over global scaling degrades as sparsity and noise increase. This is an addressable experimental gap, not a demonstrated failure, but it is load-bearing for the paper's stated motivation.
  2. [Tables 1 and 2] No repeated-sampling statistics are reported, although the sparse inputs are random. Tables 1 and 2 give single numbers for each measurement count, and at N=250 the fitted transforms can depend nontrivially on which pixels are drawn. Several comparisons that support the ordering of the methods (for example, SLF-250 versus SSF-250, or SSF-250 versus LF-LiDAR 1-beam) are separated by margins that may be within sampling variability. Reporting mean plus/minus standard deviation over at least five to ten random seeds, or paired per-image differences, would show whether the region-aware gain is systematic. Because random sampling is the only source of stochasticity in the method, this information is directly relevant to the reliability of the reported improvements.
minor comments (7)
  1. [Section 4.2, Table 1] The region-aware Median and region-aware Linear Fit baselines are introduced only by a sentence in Section 4.2, with no implementation details. Please specify how the per-region median ratio and per-region least-squares fit were computed, whether the same sparse points and the same region-expansion rule as Algorithm 1 were used, and how the results were merged.
  2. [Section 4.1] The dataset is written as "VIOD" in the text but as "VOID" in Table 2 and elsewhere; please use a single consistent spelling.
  3. [Introduction, Section 1] The sentence "As illustrated in our analysis in Figure 3" appears to refer to the patch-scale analysis shown in Figure 1, not the qualitative visualization in Figure 3. Please correct the cross-reference.
  4. [Section 4.1] The reference [22] is cited for Segment Anything, but [22] is "Segment and Recognize Anything at any granularity," whereas SAM itself is cited as [19] in Section 2.2. Please make the citation consistent.
  5. [Algorithm 1] The pseudocode mixes abstraction levels and is hard to parse as written: the region-expansion loop has no explicit termination condition, M_i^+ and D_i^+ are used before being defined, and "Step 3: Combine" appears inside the loop structure. Please rewrite the algorithm with clear definitions and a well-defined overall procedure.
  6. [Tables 1 and 2] The metric names are inconsistent: Table 1 uses "log10" while Table 2 uses "RMSElog," and the VOID table reports RMSE values in meters that are much larger than the NYUv2 RMSE values; please define all metrics and explain the difference in scale.
  7. [Reproducibility] The paper does not state whether code and random seeds will be released; for a fitting-based method with no learned parameters, releasing the code would substantially aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: sparse fitting is the stated test-time input and dense evaluation is a genuine out-of-sample measure.

full rationale

The paper's derivation chain is self-contained rather than circular. Its method is explicitly a test-time fitting procedure: sparse depth measurements are the declared input, and per-region parameters are solved only at those sparse locations. In Section 4.1 the paper states it "randomly sample[s] 250, 500, 1000, and 2000 depth points from the ground truth data," meaning the sparse points are the method's legitimate sensor-like inputs, not hidden reused labels. The reported Abs Rel, RMSE, and threshold metrics are computed over the dense ground-truth depth map; with at most 2000 of 307,200 pixels used for fitting, approximately 99% of evaluated pixels are not directly fit, so the dense error is a real generalization measure of the fitted per-region transforms. The comparison to global scaling is also not circular: global Linear Fit is a constrained special case of the region-aware model, and the paper reports empirical gains on dense evaluation rather than equating the output to the fit residual. No equation in the paper reduces to another by definition, no fitted parameter is renamed as a prediction, and no load-bearing claim rests on a self-citation; reference [11] by an author appears only as related work on window-based partitioning. The known limitations—ideal uniform sampling, absence of sensor noise, and no application of the method to simulated LiDAR beam patterns—are experimental-validity concerns about real-world robustness, not circularity in the derivation. The monotonic improvement from global to SLF to SSF is consistent with increasing model capacity, but the paper supports it with dense benchmark results on NYUv2 and VOID rather than by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method invents no physical entities. Its central claim rests on a piecewise-planar model of depth, on the assumption that SAM segmentation is depth-consistent, and on ideal sparse measurements; these are the main uncharged assumptions.

free parameters (1)
  • Per-region scale α_i, shift β_i, and plane coefficients γ_i, δ_i = Fitted per image from sparse depth samples; varies per region
    Each region's transform is estimated from the sparse points available in that region or its expansion. The number of fitted parameters grows with segment count (2 per region for SLF, 4 for SSF), so the observed improvement over a global 2-parameter fit is partly mechanical.
assumptions (3)
  • domain assumption Within each segment, metric depth is an affine function of relative depth plus a plane in image coordinates: z1 = α z2 + β x + γ y + δ.
    Invoked in Section 3.3 as the basis for SSF. It is an approximation that holds for planar indoor structure but is not derived and may fail on curved or complex geometry.
  • domain assumption Segments from SAM or OneFormer are depth-consistent: each region has a single scale and shift relationship.
    Used throughout Section 3.2 and the algorithm. SAM segments by color and texture, not depth scale, so segments may straddle depth discontinuities.
  • domain assumption Sparse depth measurements are accurate, dense enough, and pixel-aligned with the RGB image.
    Stated in Section 3.2. Experiments use random ground-truth points, so this assumption is untested for real sensors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Region-aware Depth Scale Adaptation with Sparse Measurements." pith.science (2026). https://pith.science/paper/3OAX4EFE

@misc{pith2026250714879,
  author       = {Pith},
  title        = {Pith review of: Region-aware Depth Scale Adaptation with Sparse Measurements},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3OAX4EFE}},
  note         = {Machine review of arXiv:2507.14879}
}
read the original abstract

In recent years, the emergence of foundation models for depth prediction has led to remarkable progress, particularly in zero-shot monocular depth estimation. These models generate impressive depth predictions; however, their outputs are often in relative scale rather than metric scale. This limitation poses challenges for direct deployment in real-world applications. To address this, several scale adaptation methods have been proposed to enable foundation models to produce metric depth. However, these methods are typically costly, as they require additional training on new domains and datasets. Moreover, fine-tuning these models often compromises their original generalization capabilities, limiting their adaptability across diverse scenes. In this paper, we introduce a non-learning-based approach that leverages sparse depth measurements to adapt the relative-scale predictions of foundation models into metric-scale depth. Our method requires neither retraining nor fine-tuning, thereby preserving the strong generalization ability of the original foundation models while enabling them to produce metric depth. Experimental results demonstrate the effectiveness of our approach, high-lighting its potential to bridge the gap between relative and metric depth without incurring additional computational costs or sacrificing generalization ability.

Figures

Figures reproduced from arXiv: 2507.14879 by the authors.

Figure 1
Figure 1. Analysis of Scale and Shift factor in MDE for a selected [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our proposed framework. The input image is processed by a Segmentation Foundation Model and a MDE Foundation Model, generating a segmentation map M and a relative depth prediction D. M, D are divided into multiple small regions. Within each region, sparse fitting calculations are applied to obtain a metric-scaled depth map. Finally, the metric-scaled depth maps from all regions are merged to produce the final depth … view at source ↗
Figure 3
Figure 3. Visualization of depth scale adaptation results of Depth Anything V2 on the NYU Depth V2 dataset. From left to right, the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 31 canonical work pages

  1. [1]

    Bidirectional attention network for monocular depth estimation

    Shubhra Aich, Jean Marie Uwabeza Vianney, Md Amirul Is- lam, and Mannat Kaur Bingbing Liu. Bidirectional attention network for monocular depth estimation. In 2021 IEEE In- ternational Conference on Robotics and Automation (ICRA), pages 11746–11752. IEEE, 2021. 1

    work page 2021

  2. [2]

    Adabins: Depth estimation using adaptive bins

    Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009–4018, 2021. 2, 7

    work page 2021

  3. [3]

    Zoedepth: Zero-shot trans- fer by combining relative and metric depth

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. Zoedepth: Zero-shot trans- fer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 2, 3

    arXiv 2023

  4. [4]

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

    Aleksei Bochkovskii, Ama ¨el Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 1, 3

    arXiv 2024

  5. [5]

    Ma-sam: Modality-agnostic sam adap- tation for 3d medical image segmentation

    Cheng Chen, Juzheng Miao, Dufan Wu, Aoxiao Zhong, Zhiling Yan, Sekeun Kim, Jiang Hu, Zhengliang Liu, Lichao Sun, Xiang Li, et al. Ma-sam: Modality-agnostic sam adap- tation for 3d medical image segmentation. Medical Image Analysis, 98:103310, 2024. 1, 3

    work page 2024

  6. [6]

    Sam-adapter: Adapting segment anything in underperformed scenes

    Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 3367–3375, 2023. 1, 3

    work page 2023

  7. [7]

    Katsaggelos

    Qiqin Dai, Fengqiang Li, Oliver Cossairt, and Aggelos K. Katsaggelos. Adaptive illumination based depth sensing us- ing deep superpixel and soft sampling approximation. IEEE Transactions on Computational Imaging, 8:224–235, 2022. 2

    work page 2022

  8. [8]

    Cswin transformer: A general vision transformer backbone with cross-shaped windows

    Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12124–12134, 2022. 2

    work page 2022

Show all 50 references
  1. [9]

    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. Advances in neural information processing systems , 27, 2014. 2, 7

    2014

  2. [10]

    Collaborative three-dimensional completion of color and depth in a specified area with superpixels

    Lei Fan, Long Chen, Chaoqiang Zhang, Wei Tian, and Dongpu Cao. Collaborative three-dimensional completion of color and depth in a specified area with superpixels. IEEE Transactions on Industrial Electronics , 66(8):6260–6269,

  3. [11]

    Con- trastive learning for depth prediction

    Rizhao Fan, Matteo Poggi, and Stefano Mattoccia. Con- trastive learning for depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3225–3236, 2023. 2

    2023

  4. [12]

    Deep ordinal regression net- work for monocular depth estimation

    Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat- manghelich, and Dacheng Tao. Deep ordinal regression net- work for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 2002–2011, 2018. 2

    2002

  5. [13]

    Vision meets robotics: The kitti dataset

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The Inter- national Journal of Robotics Research , 32(11):1231–1237,

  6. [14]

    Digging into self-supervised monocular depth estimation

    Cl ´ement Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 3828–3838,

  7. [15]

    Depthfm: Fast monocular depth estimation with flow matching

    Ming Gui, Johannes Schusterbauer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Bj ¨orn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 3

    2024 arXiv

  8. [16]

    Lotus: Diffusion-based visual foundation model for high-quality dense prediction

    Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Zhang, Bingbing Liu, and Ying- Cong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3

    2024 arXiv

  9. [17]

    Oneformer: One transformer to rule universal image segmentation

    Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989–2998, 2023. 2

    2023

  10. [18]

    Repurpos- ing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met- zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos- ing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9492...

    2024

  11. [19]

    Segment any- thing

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

    2023

  12. [20]

    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, 5

    1907 arXiv

  13. [21]

    Patch-wise attention network for monocular depth estimation

    Sihaeng Lee, Janghyeon Lee, Byungju Kim, Eojindl Yi, and Junmo Kim. Patch-wise attention network for monocular depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1873–1881, 2021. 1

    2021

  14. [22]

    Segment and recognize anything at any granularity

    Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, Lei Zhang, and Jianfeng Gao. Segment and recognize anything at any granularity. In Eu- ropean Conference on Computer Vision , pages 467–484. Springer, 2025. 7

    2025

  15. [23]

    Binsformer: Revisiting adaptive bins for monocular depth estimation

    Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 2

    2022 arXiv

  16. [24]

    Prompting depth anything for 4k resolution accurate metric depth estimation

    Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji- aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. 2024. 1

    2024

  17. [25]

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

    Ce Liu, Suryansh Kumar, Shuhang Gu, Radu Timofte, and Luc Van Gool. Va-depthnet: A variational approach to single image depth prediction. arXiv preprint arXiv:2302.06556 ,

    arXiv

  18. [26]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012–10022, 2021. 2

    2021

  19. [27]

    Foundation models meet low-cost sensors: Test-time adaptation for rescaling disparity for zero-shot metric depth estimation

    R ´emi Marsal, Alexandre Chapoutot, Philippe Xu, and David Filliat. Foundation models meet low-cost sensors: Test-time adaptation for rescaling disparity for zero-shot metric depth estimation. arXiv preprint arXiv:2412.14103, 2024. 2, 3, 8

    2024

  20. [28]

    Depth prompting for sensor-agnostic depth estimation

    Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9859–9869, 2024. 1, 2, 3

    2024

  21. [29]

    Unidepth: Universal monocular metric depth estimation

    Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 10106–10116. ...

    2024

  22. [30]

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

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623–1637, 2020. 2, 3, 7, 8

    2020

  23. [31]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. 8

    2021

  24. [32]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1

    2024 arXiv

  25. [33]

    Nddepth: Normal-distance as- sisted monocular depth estimation

    Shuwei Shao, Zhongcai Pei, Weihai Chen, Xingming Wu, and Zhengguo Li. Nddepth: Normal-distance as- sisted monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7931–7940. IEEE/CVF, 2023. 2

    2023

  26. [34]

    Indoor segmentation and support inference from rgbd images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European conference on computer vision , pages 746–760. Springer, 2012. 2, 5

    2012

  27. [35]

    Pdc: piecewise depth completion utilizing superpixels

    Dennis Teutscher, Patrick Mangat, and Oliver Wasenm ¨uller. Pdc: piecewise depth completion utilizing superpixels. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 2752–2758. IEEE, 2021. 2

    2021

  28. [36]

    Dcdepth: Progressive monocular depth estimation in discrete cosine domain

    Kun Wang et al. Dcdepth: Progressive monocular depth estimation in discrete cosine domain. arXiv preprint arXiv:2410.14980, 2024. Accepted by NeurIPS 2024. 2

    2024 arXiv

  29. [37]

    Unsupervised depth completion from visual iner- tial odometry

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

    1906

  30. [38]

    Toward practical monocular in- door depth estimation

    Cho-Ying Wu, Jialiang Wang, Michael Hall, Ulrich Neu- mann, and Shuochen Su. Toward practical monocular in- door depth estimation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 3814–3824, 2022. 2, 3

    2022

  31. [39]

    Medical sam adapter: Adapting seg- ment anything model for medical image segmentation.arXiv preprint arXiv:2304.12620, 2023

    Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting seg- ment anything model for medical image segmentation.arXiv preprint arXiv:2304.12620, 2023. 1, 3

    2023 arXiv

  32. [40]

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

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 1, 3, 7

    2024

  33. [41]

    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 preprint arXiv:2406.09414 , 2024. 1, 2, 3, 7

    2024 arXiv

  34. [42]

    En- forcing geometric constraints of virtual normal for depth pre- diction

    Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. En- forcing geometric constraints of virtual normal for depth pre- diction. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 5684–5693, 2019. 2

    2019

  35. [43]

    Learning to recover 3d scene shape from a single image

    Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204–213, 2021. 2

    2021

  36. [44]

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

    Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocu- lar depth estimation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 3916–3925, 2022. 7

    2022

  37. [45]

    New crfs: Neural window fully-connected crfs for monocular depth estimation

    Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. New crfs: Neural window fully-connected crfs for monocular depth estimation. arXiv preprint arXiv:2203.01502, 2022. 1, 2

    2022 arXiv

  38. [46]

    Rsa: Resolving scale ambiguities in monoc- ular depth estimators through language descriptions

    Ziyao Zeng, Yangchao Wu, Hyoungseob Park, Daniel Wang, Fengyu Yang, Stefano Soatto, Dong Lao, Byung-Woo Hong, and Alex Wong. Rsa: Resolving scale ambiguities in monoc- ular depth estimators through language descriptions. arXiv preprint arXiv:2410.02924, 2024. 1, 2, 3, 5, 8

    arXiv

  39. [47]

    Faster segment anything: Towards lightweight sam for mo- bile applications

    Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mo- bile applications. arXiv preprint arXiv:2306.14289, 2023. 1, 3

    2023 arXiv

  40. [48]

    Personalize segment anything model with one shot

    Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junt- ing Pan, Xianzheng Ma, Hao Dong, Peng Gao, and Hong- sheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 1, 3

    2023 arXiv

  41. [49]

    Metric from human: Zero-shot monoc- ular metric depth estimation via test-time adaptation

    Yizhou Zhao, Hengwei Bian, Kaihua Chen, Pengliang Ji, Liao Qu, Shao-yu Lin, Weichen Yu, Haoran Li, Hao Chen, Jun Shen, et al. Metric from human: Zero-shot monoc- ular metric depth estimation via test-time adaptation. In The Thirty-eighth Annual Conference on Neural Information...

    2024

  42. [50]

    Scaledepth: Decomposing metric depth estimation into scale prediction and relative depth estimation

    Ruijie Zhu, Chuxin Wang, Ziyang Song, Li Liu, Tianzhu Zhang, and Yongdong Zhang. Scaledepth: Decomposing metric depth estimation into scale prediction and relative depth estimation. arXiv preprint arXiv:2407.08187 , 2024. 3

    2024

Pith tools

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