Pith. sign in

REVIEW 4 major objections 4 minor 56 references

DepthDark: Robust Monocular Depth Estimation for Low-Light Environments

T0 review · 4 major / 4 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 DepthDark claims that a daylight depth foundation model fine-tuned on synthetic night-time images produced by flare and noise simulation reaches state-of-the-art accuracy on the nuScenes-Night and RobotCar-Night benchmarks.

desk verdict A practical SOTA recipe for low-light depth built from known parts, with a real gap between the claimed physical fidelity of LLDG and the evidence. read the letter →

arxiv 2507.18243 v1 pith:O4IGS52M submitted 2025-07-24 cs.CV cs.AI

Longjian Zeng , Zunjie Zhu , Rongfeng Lu , Ming Lu , Bolun Zheng , Chenggang Yan , Anke Xue This is my paper
classification cs.CVcs.AI
keywords monoculardepthestimationlow-lightsyntheticdataflaresimulationsensornoiseparameter-efficientfine-tuningilluminationguidancemultiscalefeaturefusion
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

DepthDark sets out to make monocular depth estimation reliable at night without a huge collection of real low-light depth labels. The paper argues that the two obstacles are data and adaptation: no large paired low-light RGB-D dataset exists, and directly fine-tuning a foundation model on small real night data is unstable. It proposes generating paired low-light images from ordinary daylight RGB-D data with a flare-simulation module and a noise-simulation module, then fine-tuning a pretrained depth foundation model with an illumination-guided, multiscale, parameter-efficient adapter that leaves the backbone frozen. On the nuScenes-Night and RobotCar-Night benchmarks, the model reports absolute-relative errors of 0.210 and 0.157, lower than the 0.272 and 0.235 reported for Depth Anything V2 on the same tests. If the transfer holds, the recipe turns any daylight depth foundation model into a low-light one with modest data and a few hours of single-GPU training.

What carries the argument

The load-bearing mechanism is the paired design of LLDG and LLPEFT. LLDG is a deterministic pipeline that converts a daylight RGB-D image into a low-light one by simulating the two visual signatures of night: flares and sensor noise. Flares come from a random light source sampled from the Flare7K library, positioned with a depth constraint and rendered with random darkening, gamma correction, and the Phong illumination model; noise comes from the ELD shot-read model, decomposing the total noise into photon shot, read, row, and quantization components with ISO-dependent gain. LLPEFT is the adaptation mechanism: it builds an auxiliary input by concatenating the low-light image with its channel-mean grayscale as illumination guidance, extracts features at three kernel scales, combines them with softmax attention weights, and feeds the fused features into the frozen transformer backbone through the patch embed. The key design choice is that the backbone weights stay frozen, so nearly all of the model's capacity is preserved while only a tiny plug-in is trained.

What would settle it

Train the identical LLPEFT head on a small real paired night-time RGB-D set and compare with DepthDark on nuScenes-Night and RobotCar-Night; if the real-data model does not at least match the synthetic-data model, the fidelity of LLDG to true night imaging is called into question. A cheaper check is to compare the noise variance and flare statistics of LLDG outputs against the two night benchmarks, since a large mismatch in, say, row-noise variance or flare peak intensity would identify exactly where the transfer is fragile.

Watch

Extended reading notes

Core claim

The central claim is that the low-light depth gap can be closed by separating the problem into a data-synthesis step and a parameter-efficient adaptation step. The first step, called LLDG, takes a daytime RGB-D frame, randomly darkens it, applies gamma correction, places a sampled Flare7K light source in 3D with a depth cap of 20 meters, renders it through a Phong illumination model, and adds a four-component physical noise field (photon shot, read, row, and quantization) with ISO-dependent gain following the ELD model. The second step, called LLPEFT, keeps the pretrained Depth Anything V2 backbone frozen and injects a small trainable module: it concatenates the low-light image with its channel-mean illumination guidance, runs parallel 1x1, 3x3, and 5x5 convolutions, weights the three scales with softmax attention, and feeds the result into the patch embedding. Trained only on LLDG-degraded Hypersim and Virtual KITTI pairs, the model achieves state-of-the-art numbers on nuScenes-Night and RobotCar-Night, with an ablation showing that each module contributes and that the full system beats both the base foundation model and prior domain-adaptation night-depth methods.

Load-bearing premise

The load-bearing premise is that LLDG's synthetic low-light images are statistically faithful enough to real night-time camera output that a model trained on them behaves like one trained on real night data.

Editorial extensions

If this is right

  • If DepthDark is correct, large real low-light RGB-D datasets are not a prerequisite for night-time monocular depth; degrading existing day datasets with flare and noise simulation can substitute.
  • A foundation model adapted this way transfers to night driving benchmarks it never trained on, which is the generalization behavior a robust low-light model needs.
  • The parameter-efficient adapter adds only about 9,000 parameters over the base model (97.479M versus 97.470M), so the same fine-tuning recipe is cheap enough to repeat for other backbones or image-degradation conditions.
  • The ablations in the paper show each component independently improves on the base model, and the two together give the best results, meaning the synthesis and adaptation pieces are complementary rather than redundant.

Reading between the lines

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

  • Editorial inference: the cleanest test of the synthetic-data claim is to fine-tune the same LLPEFT plug-in on a small set of real paired night images; if that reaches or beats DepthDark's numbers, the remaining bottleneck is indeed data distribution rather than the adapter.
  • Editorial inference: the illumination guidance is only a channel-mean grayscale, so a natural family of extensions would swap in richer guidance such as denoised luminance, local exposure maps, or a learned illumination estimator without changing the architecture.
  • Editorial inference: because the noise parameters follow a specific ISO/gain calibration, the reported gains may be sensitive to camera and ISP; calibrating the noise model on a target sensor is a direct way to test how far the transfer generalizes.
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 / 4 minor

Summary. The paper proposes DepthDark, a monocular depth estimation foundation model for low-light scenes. It introduces a Low-Light Dataset Generation (LLDG) pipeline that synthesizes paired low-light RGB-D training data from daytime images using a flare-simulation module (FSM) based on Flare7K light sources and a Phong illumination model, plus a noise-simulation module (NSM) based on a shot-read noise model. It also proposes a Low-Light Parameter-Efficient Fine-Tuning (LLPEFT) strategy that concatenates an illumination-guidance grayscale image with the low-light input and fuses multiscale features before feeding the DPT decoder of Depth Anything V2. The model is trained on Hypersim and Virtual KITTI and evaluated on nuScenes-Night and RobotCar-Night, reporting ABS rel 0.210 and 0.157 respectively, which the paper claims is state-of-the-art. Ablation studies decompose the contribution of LLDG and LLPEFT and compare LLPEFT against AMFG and LoRA.

Significance. If the reported results hold, the paper makes a practically useful contribution: it shows that a strong depth foundation model can be adapted to low-light conditions with a relatively small synthetic dataset and limited compute, which lowers the barrier to low-light depth estimation. The evaluation setup is largely sound in that the held-out nuScenes-Night and RobotCar-Night tests are not used to generate the synthetic training data, and the comparison set includes current state-of-the-art low-light and foundation-model baselines. The ablation that separates LLDG from LLPEFT is a good design choice. However, the significance is materially limited by the lack of any validation that the synthesized degradation matches real nighttime image statistics, by the absence of a generic-degradation control baseline, and by the absence of released code or data. The paper's generalization claim is also weakened by an internal contradiction about whether the training data contains autonomous-driving samples.

major comments (4)
  1. [§3.1, Tables 4.2–4.3] The central claim that the physically motivated flare and noise modules in LLDG are responsible for the reported gains is not supported by the experiments as presented. Tables 4.2 and 4.3 compare the full LLDG+LLPEFT pipeline against variants that remove one component, but there is no control condition in which the same daytime images are degraded by a simple baseline such as random darkening, gamma correction, and additive Gaussian or Poisson noise. Without such a control, the results are consistent with the alternative hypothesis that most of the improvement comes from fine-tuning a strong DINOv2/Depth Anything V2 backbone on any low-light-like augmentation, and the specific flare/noise design remains unverified. I ask the authors to add this control and, ideally, a quantitative comparison of noise and flare statistics between LLDG outputs and real nuScenes-Night/RobotCar-Night images.
  2. [§3.1.2, Eq. (6)] The noise model is taken from ELD [37], which is a RAW-domain model: photon shot noise, read noise, row noise, and quantization noise are defined on sensor counts before demosaicing and tone mapping. Here the model is applied to I_F, which has already undergone gamma correction and brightness scaling in Eq. (4). Applying K*N_p + N_read + N_r + N_q to a gamma-encoded, sRGB-like image is physically inconsistent and makes the effective noise levels unclear. The authors should either inject noise in the linear domain before the darkening and gamma operations, or provide evidence that the sRGB-domain approximation is an adequate model for the target cameras in nuScenes-Night and RobotCar-Night.
  3. [§4.3.1 vs. §4.2.1] The text in §4.3.1 states that 'our training data contained no samples from the autonomous driving domain' and uses this claim to argue for generalization to nuScenes-Night and RobotCar-Night. This is contradicted by §4.2.1, which lists Virtual KITTI [4] as a training dataset; Virtual KITTI is a synthetic autonomous-driving dataset. This factual inconsistency undermines the generalization interpretation. The authors should correct the claim, for example by stating that no real driving data was used, or ideally retrain without Virtual KITTI to verify that the transfer result holds for truly non-driving training data.
  4. [Tables 4.2 and 4.3] The row labeled 'Depth Anything V2' in Table 4.3 has exactly the same reported values as the '+ Only LLDG' row in Table 4.2 on both benchmarks (e.g., nuScenes: ABS rel 0.264, Sq rel 2.956, RMSE 9.209; RobotCar: ABS rel 0.183, Sq rel 1.679, RMSE 6.769). If this row denotes full fine-tuning on the LLDG dataset, that should be stated explicitly in the table or caption. As written, the reader cannot tell whether the table reports the original untuned checkpoint or a fine-tuned one, which makes the parameter-count comparison in the PEFT ablation difficult to interpret.
minor comments (4)
  1. [§3.2.1, Eq. (7)] The notation LLDG(I) in Eq. (7) is ambiguous because LLDG is introduced as a dataset-generation framework rather than an image-to-image function. Define I_FN = LLDG(I) explicitly at first use to avoid confusion.
  2. [§4.1] The sentence 'we adopted the training settings from Depth Anything V2 and used the DPT decoder [30] in the ZoeDepth [3] pipeline' should specify which components are frozen, which are trainable, and where the LLPEFT modules are inserted in the network. This information is needed to understand the parameter counts in Table 4.3.
  3. [Figures 4.1 and 4.2] The captions mention 'N-N & R-N' but the running text refers to the nuScenes-Night and RobotCar-Night datasets. The captions should name each dataset explicitly so the reader does not have to infer the abbreviations.
  4. [General] The paper does not state whether code or the LLDG-generated training dataset will be made available. Since one of the main contributions is the synthetic dataset-generation pipeline, adding a release statement or a clear explanation of why the assets are withheld would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: evaluation is on held-out real benchmarks and no prediction reduces to a fitted input.

full rationale

The paper's central derivation is self-contained and externally validated. DepthDark is trained on synthetic low-light data produced by LLDG and fine-tuned with LLPEFT, and it is evaluated on held-out real benchmark datasets nuScenes-Night and RobotCar-Night, with comparison numbers taken from external prior work. No parameter is fitted to the test sets, and no test-set quantity is renamed as a prediction. The ablation studies in Tables 4.2 and 4.3 show each proposed module improves results on those same external benchmarks, so the claimed contributions are not equivalent to their inputs by construction. The noise and flare models are imported from external sources (ELD, Flare7K, and standard shot-read noise formulations), not from the authors' own prior work, and the paper does not invoke any self-citation chain to force its design choices. Concerns that LLDG's synthetic distribution is not quantitatively validated against real night statistics are correctness or generalization risks, not circularity, because the evaluation criterion is external to the data-generation pipeline.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-chosen synthesis distributions (flare intensity, brightness, gamma, noise parameters) and several domain assumptions about how well synthetic degradation transfers to real night cameras. No new physical entity is introduced. Learned neural-network weights are not listed as free parameters; the listed items are the manually set quantities that shape the synthetic training distribution.

free parameters (7)
  • Light source depth upper limit = 20 m
    Hand-chosen bound in Eq. (1) that prevents distant light sources from producing excessive flare artifacts in FSM.
  • Brightness scaling factor s_b = sampled from U(0.4, 1)
    Hand-chosen random darkening range applied to daylight images before flare synthesis in Eq. (4).
  • Gamma correction factor g_F = sampled from U(1.8, 2.2)
    Hand-chosen range controlling the low-light tone mapping in Eq. (4).
  • Flare scale factor s_F distribution = log-uniform with hand-set min/max
    Controls per-light-source strength in Eq. (2); the bounds are not reported in the paper.
  • Flare peak intensity F distribution = log-uniform with hand-set min/max
    Controls total flare intensity and the number of light sources in Eqs. (2) and (3); the bounds are not reported.
  • Noise model parameters from ELD = ISO-dependent gain K and variances
    Shot, read, row, and quantization noise parameters are assumed from ELD [37] for the synthetic training images in Eqs. (5) and (6).
  • Multiscale fusion output channels C3 = 32 (from figure Conv k1c32)
    Hand-chosen channel count in the multiscale feature fusion branch of LLPEFT.
assumptions (5)
  • domain assumption The Phong illumination model combined with random darkening and gamma correction adequately approximates nighttime flare and brightness distribution.
    Used in Eq. (4) to produce synthetic flare images; no quantitative validation against real night flare statistics is provided.
  • domain assumption The ELD shot-read noise decomposition describes real low-light camera noise for the target benchmarks.
    Invoked in Section 3.1.2, Eqs. (5) and (6); parameters are taken from ELD [37] and not measured for nuScenes or RobotCar cameras.
  • domain assumption The Flare7K light source library contains flare patterns representative of the target night scenes.
    Sampled in FSM in Section 3.1.1 to simulate glare; the paper provides no evaluation of the distribution gap.
  • domain assumption Depth Anything V2 is a suitable pretrained foundation model for low-light fine-tuning with a DPT decoder.
    Used as the backbone throughout the paper; no comparison with other backbones is reported.
  • domain assumption Hypersim and Virtual KITTI depth labels are accurate and diverse enough to support low-light transfer.
    Used as the source of paired depth for LLDG in Section 4.2.1; the paper says KITTI labels are unreliable but does not separately validate the chosen datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DepthDark: Robust Monocular Depth Estimation for Low-Light Environments." pith.science (2026). https://pith.science/paper/O4IGS52M

@misc{pith2026250718243,
  author       = {Pith},
  title        = {Pith review of: DepthDark: Robust Monocular Depth Estimation for Low-Light Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O4IGS52M}},
  note         = {Machine review of arXiv:2507.18243}
}
read the original abstract

In recent years, foundation models for monocular depth estimation have received increasing attention. Current methods mainly address typical daylight conditions, but their effectiveness notably decreases in low-light environments. There is a lack of robust foundational models for monocular depth estimation specifically designed for low-light scenarios. This largely stems from the absence of large-scale, high-quality paired depth datasets for low-light conditions and the effective parameter-efficient fine-tuning (PEFT) strategy. To address these challenges, we propose DepthDark, a robust foundation model for low-light monocular depth estimation. We first introduce a flare-simulation module and a noise-simulation module to accurately simulate the imaging process under nighttime conditions, producing high-quality paired depth datasets for low-light conditions. Additionally, we present an effective low-light PEFT strategy that utilizes illumination guidance and multiscale feature fusion to enhance the model's capability in low-light environments. Our method achieves state-of-the-art depth estimation performance on the challenging nuScenes-Night and RobotCar-Night datasets, validating its effectiveness using limited training data and computing resources.

Figures

Figures reproduced from arXiv: 2507.18243 by the authors.

Figure 1.1
Figure 1.1. Overview of our DepthDark Training Framework. The framework consists of LLDG and LLPEFT. The Patch Embed [PITH_FULL_IMAGE:figures/full_fig_p003_1_1.png] view at source ↗
Figure 3.1
Figure 3.1. Paired visual examples for different scenes, [PITH_FULL_IMAGE:figures/full_fig_p004_3_1.png] view at source ↗
Figure 4.1
Figure 4.1. Qualitative comparison results of different monocular depth estimation methods on the nuScenes-Night dataset. For [PITH_FULL_IMAGE:figures/full_fig_p007_4_1.png] view at source ↗
Figures from the paper (1 more)
Figure 4.2
Figure 4.2. Figure 4.2: Qualitative comparison results of different monocular depth estimation methods on the RobotCar-Night dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_4_2.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 42 canonical work pages

  1. [37]

    Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. 2020. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 2758–2767

    work page 2020

  2. [4]

    Yohann Cabon, Naila Murray, and Martin Humenberger. 2020. Virtual kitti 2. arXiv preprint arXiv:2001.10773 (2020)

    arXiv 2020

  3. [1]

    Jinwoo Bae, Sungho Moon, and Sunghoon Im. 2023. Deep digging into the generalization of self-supervised monocular depth estimation. In Proceedings of the AAAI conference on artificial intelligence , Vol. 37. 187–196

    work page 2023

  4. [2]

    Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. 2022. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274 (2022)

    arXiv 2022

  5. [3]

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller

  6. [5]

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. 2023. Retinexformer: One-stage retinex-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 12504–12513

    work page 2023

  7. [6]

    Wei-Ting Chen, Yu-Jiet Vong, Sy-Yen Kuo, Sizhou Ma, and Jian Wang. 2024. RobustSAM: Segment Anything Robustly on Degraded Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 4081–4091

    work page 2024

  8. [7]

    Ri Cheng, Yuqi Sun, Bo Yan, Weimin Tan, and Chenxi Ma. 2022. Geometry-aware reference synthesis for multi-view image super-resolution. In Proceedings of the 30th ACM International Conference on Multimedia . 6083–6093

    work page 2022

Show all 56 references
  1. [8]

    Yuekun Dai, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy

  2. [9]

    Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  3. [10]

    Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattad. 2023. Generative models: What do they know? do they know things? let’s find out! arXiv preprint arXiv:2311.17137 (2023)

    2023 arXiv

  4. [11]

    David Eigen, Christian Puhrsch, and Rob Fergus. 2014. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014)

    2014

  5. [12]

    Hansen Feng, Lizhi Wang, Yuzhi Wang, and Hua Huang. 2022. Learnability enhancement for low-light raw denoising: Where paired real data meets noise modeling. In Proceedings of the 30th ACM International Conference on Multimedia . 1436–1444

    2022

  6. [13]

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. 2024. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision 132, 2 (2024), 581–595

    2024

  7. [14]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

    2016

  8. [15]

    Derek Hoiem, Alexei A Efros, and Martial Hebert. 2007. Recovering surface layout from an image. International Journal of Computer Vision 75 (2007), 151–172

    2007

  9. [16]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

    2021 arXiv

  10. [17]

    Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. 2021. Neigh- bor2neighbor: Self-supervised denoising from single noisy images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 14781– 14790

    2021

  11. [18]

    Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. 2022. Neigh- bor2Neighbor: A self-supervised framework for deep image denoising. IEEE Transactions on Image Processing 31 (2022), 4023–4038

    2022

  12. [19]

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. 2024. Repurposing diffusion-based image genera- tors for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 9492–9502

    2024

  13. [20]

    Bo Li, Chunhua Shen, Yuchao Dai, Anton Van Den Hengel, and Mingyi He. 2015. Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In Proceedings of the IEEE conference on computer vision and pattern recognition . 1119–1127

    2015

  14. [21]

    Lifeng Lin, Rongfeng Lu, Quan Chen, Haofan Ren, Ming Lu, Yaoqi Sun, Cheng- gang Yan, and Anke Xue. 2025. VGNC: Reducing the Overfitting of Sparse- view 3DGS via Validation-guided Gaussian Number Control. arXiv preprint arXiv:2504.14548 (2025)

    2025 arXiv

  15. [22]

    Ariel Lipson, Stephen G Lipson, and Henry Lipson. 2010. Optical physics. Cam- bridge University Press

    2010

  16. [23]

    Ce Liu, Jenny Yuen, Antonio Torralba, Josef Sivic, and William T Freeman. 2008. Sift flow: Dense correspondence across different scenes. In Computer Vision– ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, Octo- ber 12-18, 2008, Proceedings, Part III ...

    2008

  17. [24]

    Lina Liu, Xibin Song, Mengmeng Wang, Yong Liu, and Liangjun Zhang. 2021. Self-supervised monocular depth estimation for all day images using domain separation. In Proceedings of the IEEE/CVF international conference on computer vision. 12737–12746

    2021

  18. [25]

    Rongfeng Lu, Hangyu Chen, Zunjie Zhu, Yuhang Qin, Ming Lu, Le Zhang, Cheng- gang Yan, and Anke Xue. 2024. Thermalgaussian: Thermal 3d gaussian splatting. arXiv preprint arXiv:2409.07200 (2024)

    2024 arXiv

  19. [26]

    Rongfeng Lu, Zunjie Zhu, Sheng Fu, Shenrong Chen, Tingyu Wang, Chenggang Yan, and Feng Xu. 2023. Self-supervised camera relocalization with hierarchical fern encoding. IEEE Transactions on Instrumentation and Measurement 73 (2023), 1–12

    2023

  20. [27]

    YiFan Lu, Ning Xie, and Heng Tao Shen. 2020. DMCR-GAN: adversarial denoising for monte carlo renderings with residual attention networks and hierarchical features modulation of auxiliary buffers. In SIGGRAPH Asia 2020 Technical Communications. 1–4

    2020

  21. [28]

    Bui Tuong Phong. 1998. Illumination for computer generated pictures. InSeminal graphics: pioneering efforts that shaped the field . 95–101

    1998

  22. [29]

    Weichao Qiu, Fangwei Zhong, Yi Zhang, Siyuan Qiao, Zihao Xiao, Tae Soo Kim, and Yizhou Wang. 2017. Unrealcv: Virtual worlds for computer vision. In Proceedings of the 25th ACM international conference on Multimedia . 1221–1224

    2017

  23. [30]

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. 2021. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision. 12179–12188

    2021

  24. [31]

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. 2021. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conferen...

    2021

  25. [32]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  26. [33]

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

    2023

  27. [34]

    Madhu Vankadari, Sourav Garg, Anima Majumder, Swagat Kumar, and Ardhendu Behera. 2020. Unsupervised monocular depth estimation for night-time images using adversarial domain feature adaptation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, ...

    2020

  28. [35]

    Madhu Vankadari, Stuart Golodetz, Sourav Garg, Sangyun Shin, Andrew Markham, and Niki Trigoni. 2023. When the sun goes down: Repairing pho- tometric losses for all-day depth estimation. In Conference on Robot Learning . PMLR, 1992–2003

    2023

  29. [36]

    Kun Wang, Zhenyu Zhang, Zhiqiang Yan, Xiang Li, Baobei Xu, Jun Li, and Jian Yang. 2021. Regularizing nighttime weirdness: Efficient self-supervised monoc- ular depth estimation in the dark. In Proceedings of the IEEE/CVF international conference on computer vision . 16055–16064

    2021

  30. [38]

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

    2020

  31. [39]

    Zhongwei Xuan, Zunjie Zhu, Shuai Wang, Haibing Yin, Hongkui Wang, and Ming Lu. 2024. Superpixel-based Efficient Sampling for Learning Neural Fields from Large Input. In Proceedings of the 32nd ACM International Conference on Multimedia. 10421–10430

    2024

  32. [40]

    Haolin Yang, Chaoqiang Zhao, Lu Sheng, and Yang Tang. 2024. Self-Supervised Monocular Depth Estimation in the Dark: Towards Data Distribution Compensa- tion. arXiv preprint arXiv:2404.13854 (2024)

    2024 arXiv

  33. [41]

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

    2024

  34. [42]

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. 2024. Depth Anything V2. arXiv preprint arXiv:2406.09414 (2024). MM ’25, October 27–31, 2025, Dublin, Ireland. Longjian Zeng et al

    2024 arXiv

  35. [43]

    Xiaodong Yang, Zhuang Ma, Zhiyu Ji, and Zhe Ren. 2023. Gedepth: Ground embedding for monocular depth estimation. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision . 12719–12727

    2023

  36. [44]

    Hantao Yao, Rui Zhang, and Changsheng Xu. 2023. Visual-language prompt tun- ing with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 6757–6767

    2023

  37. [45]

    Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. 2019. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF international conference on computer vision . 5684–5693

    2019

  38. [46]

    Jiaqi Yu, Yongwei Nie, Chengjiang Long, Wenjun Xu, Qing Zhang, and Guiqing Li. 2021. Monte Carlo denoising via auxiliary feature guided self-attention. ACM Trans. Graph. 40, 6 (2021), 273–1

    2021

  39. [47]

    Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. 2023. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...

    2023

  40. [48]

    Yi Zhang, Hongwei Qin, Xiaogang Wang, and Hongsheng Li. 2021. Rethinking noise synthesis and modeling in raw denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 4593–4601

    2021

  41. [49]

    Zhedong Zhang, Liang Li, Gaoxiang Cong, Haibing Yin, Yuhan Gao, Chenggang Yan, Anton van den Hengel, and Yuankai Qi. 2024. From speaker to dubber: movie dubbing with prosody and duration consistency learning. In Proceedings of the 32nd ACM International Conference on Multimedi...

    2024

  42. [50]

    Chaoqiang Zhao, Yang Tang, and Qiyu Sun. 2022. Unsupervised monocular depth estimation in highly complex environments.IEEE Transactions on Emerging Topics in Computational Intelligence 6, 5 (2022), 1237–1246

    2022

  43. [51]

    Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. 2022. Monovit: Self- supervised monocular depth estimation with a vision transformer. In 2022 inter- national conference on 3D vision (3DV) . IEEE, 668–678

    2022

  44. [52]

    Hengrun Zhao, Bolun Zheng, Shanxin Yuan, Hua Zhang, Chenggang Yan, Liang Li, and Gregory Slabaugh. 2022. CBREN: Convolutional Neural Networks for Constant Bit Rate Video Quality Enhancement. IEEE Transactions on Circuits and Systems for Video Technology 32, 7 (2022), 4138–4149

    2022

  45. [53]

    Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. 2023. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 5729– 5739

    2023

  46. [54]

    B Zheng, S Yuan, C Yan, X Tian, J Zhang, Y Sun, L Liu, A Leonardis, and G Slabaugh. 2022. Learning Frequency Domain Priors for Image Demoireing. IEEE Transactions on Pattern Analysis and Machine Intelligence44, 11 (2022), 7705–7717

    2022

  47. [2022]

    Advances in Neural Information Processing Systems 35 (2022), 3926–3937

    Flare7k: A phenomenological nighttime flare removal dataset. Advances in Neural Information Processing Systems 35 (2022), 3926–3937

    2022

  48. [2023]

    Zoedepth: Zero-shot transfer by combining relative and metric depth.arXiv preprint arXiv:2302.12288 (2023)

    2023 arXiv

Pith tools

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