Pith. sign in

REVIEW 3 major objections 5 minor 68 references

RAW Image Reconstruction from RGB on Smartphones. NTIRE 2025 Challenge Report

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Smartphone RGB-to-RAW reconstruction without metadata reaches 27.66 dB PSNR in a public challenge, and small efficient models generalize best to unseen phone sensors.

desk verdict The new smartphone sRGB-RAW dataset with an out-of-device split is genuinely useful, but the paper's own tables undermine its headline SOTA and generalization claims. read the letter →

arxiv 2506.01947 v1 pith:GKONGU45 submitted 2025-06-02 eess.IV cs.CV

Marcos V. Conde , Radu Timofte , Radu Berdan , Beril Besbinar , Daisuke Iso , Pengzhou Ji , Xiong Dun , Zeying Fan
show 25 more authors
Chen Wu Zhansheng Wang Pengbo Zhang Jiazi Huang Qinglin Liu Wei Yu Shengping Zhang Xiangyang Ji Kyungsik Kim Minkyung Kim Hwalmin Lee Hekun Ma Huan Zheng Yanyan Wei Zhao Zhang Jing Fang Meilin Gao Xiang Yu Shangbin Xie Mengyuan Sun Huanjing Yue Jingyu Yang Huize Cheng Shaomeng Zhang Zhaoyang Zhang Haoxiang Liang
This is my paper · ORCID
classification eess.IVcs.CV
keywords RAWreconstructionreverseISPsRGBtosmartphoneimagingimagesignalprocessordeeplearningbenchmarkgeneralizationunseensensorsNTIREchallenge
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

This paper reports the NTIRE 2025 challenge on reconstructing RAW sensor images from ordinary smartphone sRGB photos without any metadata. The organizers built a public dataset of real RAW–RGB pairs from four phone models, trained participants on two of them, and tested on the other two unseen models. They claim the resulting benchmark establishes the state of the art for generating realistic RAW data, with the winning model, DBNet, reaching 27.66 dB PSNR and 0.770 SSIM overall. The paper also reports that simple, parameter-efficient models generalized better to unseen sensors than large ones, suggesting that heavy models overfit the training devices.

What carries the argument

The load-bearing object is the benchmark itself: a public dataset of real smartphone RAW–RGB pairs, standardized to an RGGB Bayer pattern with white and black levels corrected, split into 1,446 training pairs from an iPhone X and a Samsung S9, plus a test set of 120 images from those same target devices and 60 images from unseen Samsung S21 and Vivo X90 sensors. Test-time metadata is withheld, so models must infer ISP behavior purely from pixels. The evaluation computes PSNR and SSIM directly in the 12-bit RAW domain and separates submissions into an efficient track (under 0.2M parameters, able to process 12MP images) and a general track. Among the submitted architectures, DBNet's dynamic-bias convolution, which adapts the convolution bias to the input image at negligible parameter cost, is the mechanism the winning team credits for balancing lightness and fitting ability.

What would settle it

Take the same trained models and test them on a set of RAW–RGB pairs that includes very dark and overexposed scenes and more than two unseen phone models, ideally from different sensor generations; if the efficient models' out-of-factory advantage shrinks or the ranking changes materially, the benchmark's generalization claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a metadata-free, learning-based reverse-ISP benchmark on smartphone images is now feasible and that it advances the state of the art for realistic RAW generation. Reversing the camera's image signal processor without access to white-balance gains, color correction matrices, or other metadata is hard, especially for smartphone ISPs, which are more complex than DSLR pipelines. The challenge shows that the best submitted network—DBNet, a dynamic-bias convolution network under 0.2M parameters—achieves the top overall fidelity (27.66 dB PSNR, 0.770 SSIM) on a held-out test set of 120 images from known devices and 60 images from two unseen phone models. The paper's additional, perhaps more general, finding is that the efficient track's simple models generalized to out-of-factory devices better than the larger general-track models, meaning the benchmark's ranking favors methods that do not overfit the training sensors.

Load-bearing premise

The evaluation assumes that PSNR/SSIM computed in the RAW domain on the challenge's 180-image test set, whose out-of-factory devices are just two phone models, is a sufficient and unbiased measure of 'realistic RAW data' generation.

Editorial extensions

If this is right

  • Metadata-free RGB-to-RAW reconstruction on phones is practically achievable: the winning model hits 27.66 dB PSNR overall, and even the strongest baseline exceeds 26 dB.
  • Efficient models under 0.2M parameters can beat much larger networks on unseen sensors, implying that overfitting to known devices is the main obstacle to generalization in reverse ISP.
  • The public dataset gives future researchers a fixed, reproducible test bed for comparing reverse-ISP methods on smartphone imagery.
  • If the benchmark is accepted, it replaces the previous reference point for RGB-to-RAW reconstruction, with DBNet as the new state of the art and a clear efficiency advantage.

Reading between the lines

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

  • If the observed generalization trend holds beyond these two unseen phones, deploying reverse ISP on a new smartphone model might only require a small fine-tuning set from that sensor rather than retraining from scratch.
  • The RAW-domain PSNR/SSIM metric does not directly measure whether reconstructed RAW is 'realistic' for downstream tasks; a test that feeds reconstructed RAW into a denoiser or detector trained on true RAW would be a stricter check of the benchmark's claim.
  • Because the dataset filters out extremely dark and overexposed images, the ranking likely understates how hard real-world extreme lighting is; a stress-test set with those frames could reveal larger gaps between methods.
  • The efficiency advantage of simple models may partly reflect that the test-time data distribution (same four sensors, similar scenes) is close to training; with more diverse out-of-factory sensors, larger models might close the gap.
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

3 major / 5 minor

Summary. This paper is the NTIRE 2025 challenge report on reconstructing RAW sensor images from smartphone sRGB images without using metadata. It introduces a dataset of paired RAW-RGB images from four smartphones, defines a general track and an efficient track (limited to 0.2M parameters), reports benchmark results in Table 1, and provides descriptions of the top submitted methods. The paper claims that the benchmark establishes the state of the art for realistic RAW generation and that only simple and efficient methods generalize well to unknown out-of-domain (OOF) devices.

Significance. If the reported results are internally consistent, the paper provides a public dataset and a reproducible ranking for RGB-to-RAW reconstruction on smartphone imagery, which is a useful resource for the low-level vision community. The inclusion of both target-device and OOF-device evaluation is a valuable feature, and the paper makes concrete technical contributions from several teams (e.g., DualRAW, DBNet, ULite). However, the central claims are currently compromised by inconsistencies between the track definitions, the benchmark table, and the technical summary table, so the stated state-of-the-art and generalization conclusions are not yet supported.

major comments (3)
  1. [§2.2, §2.3, Table 1, Table 7] DBNet is listed in the efficient track in Table 1, but Table 7 reports that DBNet has 0.39M parameters, which exceeds the 0.2M parameter limit for the efficient track stated in Section 2.2. Since Section 2.3 highlights DBNet as the best proposed method, the efficient-track ranking and the headline result are not supported as stated; the authors must either correct the parameter count, move DBNet to the general track, or revise the track definition.
  2. [§2.3, Table 1] The claim that 'only the simple and efficient methods avoid overfitting and generalize on unknown OOF devices' is contradicted by the same table: ResUNet, a general-track model with 5M parameters, achieves the highest OOF PSNR (25.17 dB) and OOF SSIM (0.7196), while the efficient-track methods DBNet (23.94 dB), ULite (22.15 dB), and GAR2Net (20.74 dB) are all lower. The stated generalization conclusion cannot be drawn from the reported data without additional analysis or a clarified definition of 'simple'.
  3. [§3.10, Table 6] Res-CSP reports 29.78 dB PSNR and 0.92 SSIM on the RGB2RAW Target test set, which would be the best result in Table 1, yet this method is absent from the official benchmark table and its results are not included in the overall comparison. If these numbers were obtained under the same evaluation protocol as Table 1, the leaderboard is incomplete; if not, the section should explicitly describe the differences in test data, preprocessing, or metric computation.
minor comments (5)
  1. [Title] The title contains a spurious space: 'RA W Image Reconstruction' should be 'RAW Image Reconstruction'.
  2. [§3.7 vs Table 7] Section 3.7 states that ResUNet has 4M parameters, while Table 7 lists 5M; this discrepancy should be resolved.
  3. [Table 7] Table 7 omits ULite, UNAFNet, and Res-CSP even though these methods are described in Sections 3.4, 3.8, and 3.10; including them would make the technical summary complete.
  4. [§2.1] The phrase 'already white-black level corrected' is unclear; it presumably means 'white and black level corrected' and should be rephrased.
  5. [Figure 3 caption] The caption contains a typo: 'Overiew' should be 'Overview'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SOTA and generalization claims are held-out empirical findings; the noted inconsistencies are correctness risks, not circular reasoning.

full rationale

This paper is a challenge report whose central claims—that the benchmark establishes a state-of-the-art ranking and that simple efficient models generalize to out-of-distribution (OOF) devices—are empirical statements evaluated on a held-out test set (120 target plus 60 OOF images) described in Secs. 2.1 and 2.3. No equation in the manuscript constructs a predicted quantity from the same fitted parameter that it is supposed to predict; ULite's learned transformation matrix, DBNet's dynamic-bias convolution, and TDMFNet's gamma-path fusion are all network designs trained with supervised losses on the provided training data. The self-citations that appear are not load-bearing: ReRAW [2] is introduced as a baseline in Sec. 2.2 (“ReRAW [2] represents the state-of-the-art on RAW image reconstruction for DSLR and DSLM cameras”) and then independently evaluated on the challenge test set, where it ranks below several submitted methods; the earlier AIM challenge [14] is cited only as lineage. No uniqueness theorem or ansatz is imported from the authors' prior work to forbid alternative approaches. The reviewer-noted problems—DBNet's 0.39M parameter count conflicting with the stated 0.2M efficient-track cap (Table 1 versus Table 7) and ResUNet having the highest OOF PSNR despite the “simple and efficient” generalization statement—are internal inconsistencies or correctness risks, not circularity, because the benchmark result does not reduce to its own inputs by construction. Under the hard rule that only quoted, specific reductions count as circularity, the honest finding is no significant circularity; this benchmark is self-contained against its own held-out evaluation.

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

No derivation requiring fitted constants is present. The central claims rest on dataset and metric assumptions, listed above as domain assumptions and an ad-hoc preprocessing choice.

assumptions (3)
  • domain assumption PSNR/SSIM in the RAW domain measure the quality of reconstructed RAW images.
    The paper uses PSNR and SSIM as the sole evaluation metrics in Table 1 and equates high values with 'realistic RAW data'.
  • domain assumption The out-of-device test images from Samsung S21 and Vivo X90 represent unknown smartphone sensors.
    The paper uses these two devices to test generalization to unseen hardware (Sec. 2.1) but does not justify their representativeness.
  • ad hoc to paper Manual filtering of the dataset does not bias the benchmark toward easy cases.
    Sec. 2.1 states that extremely dark or overexposed images are manually removed, which may make the reconstruction task easier than real-world conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RAW Image Reconstruction from RGB on Smartphones. NTIRE 2025 Challenge Report." pith.science (2026). https://pith.science/paper/GKONGU45

@misc{pith2026250601947,
  author       = {Pith},
  title        = {Pith review of: RAW Image Reconstruction from RGB on Smartphones. NTIRE 2025 Challenge Report},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GKONGU45}},
  note         = {Machine review of arXiv:2506.01947}
}
read the original abstract

Numerous low-level vision tasks operate in the RAW domain due to its linear properties, bit depth, and sensor designs. Despite this, RAW image datasets are scarce and more expensive to collect than the already large and public sRGB datasets. For this reason, many approaches try to generate realistic RAW images using sensor information and sRGB images. This paper covers the second challenge on RAW Reconstruction from sRGB (Reverse ISP). We aim to recover RAW sensor images from smartphones given the corresponding sRGB images without metadata and, by doing this, ``reverse" the ISP transformation. Over 150 participants joined this NTIRE 2025 challenge and submitted efficient models. The proposed methods and benchmark establish the state-of-the-art for generating realistic RAW data.

Figures

Figures reproduced from arXiv: 2506.01947 by the authors.

Figure 1
Figure 1. DualRAW Architecture: A Uformer-based Global Branch extracts global features from the input RGB image. Parallel UNET [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Illustration of ReRAW [2] architecture and training data flow. A Global Context Encoder (GCE) extracts features from the full RGB image to guide the Color Reconstruction network (CRN), while a Multi-head Gamma Predictor (MGP) generates multiple gamma￾corrected RAW patches. These patches are then de-gammaed (inverse gamma correction), scaled by a scaling vector, predicted by a Gamma Scaling Encoder (GSE) from the ori… view at source ↗
Figure 3
Figure 3. Team TongJi-IPOE. Overiew of the proposed DBNet. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Overview of our ULite architecture proposed [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Architecture of GAR2Net by team IVISLAB. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Overall architecture of the proposed network for Reverse ISP by Team VIP. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Structure of the Residual Attention Block [ [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: The architecture of our ResUNet for RAW image recon [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: (a) Overview of our proposed network architecture. The [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: (a) Simplified Channel Attention (SCA), and (b) Sim [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Overview of the proposed TDMFNet. When calculating Lp, we average the G1 and G2 chan￾nels of RAW images to match the input channel number of pre-trained model. Experimental results demonstrate that the perceptual loss effectively suppresses the lateral arti￾facts caus…
Figure 12
Figure 12. Figure 12: Res-CSP Network: combines the benefits of ResNet with the feature extraction capabilities of the attention mechanism. CSP modules to efficiently process input images.The core network structure is shown in [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 29 canonical work pages

  1. [1]

    Semi-supervised raw-to-raw mapping

    Mahmoud Afifi and Abdullah Abuolaim. Semi-supervised raw-to-raw mapping. InBritish Machine Vision Conference (BMVC), 2021. 2

    work page 2021

  2. [2]

    ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge

    Radu Berdan, Beril Besbinar, Christoph Reinders, Junji Ot- suka, and Daisuke Iso. Reraw: Rgb-to-raw image recon- struction via stratified sampling for efficient object detection on the edge.arXiv preprint arXiv:2503.03782, 2025. 1, 2, 3, 4, 5, 12

    work page Pith review arXiv 2025

  3. [3]

    E Bousias Alexakis and C Armenakis. Evaluation of unet and unet++ architectures in high resolution image change de- tection applications.The International Archives of the Pho- togrammetry, Remote Sensing and Spatial Information Sci- ences, 43:1507–1514, 2020. 12

    work page 2020

  4. [4]

    Unprocessing images for learned raw denoising

    Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11036–11045, 2019. 1

    2019

  5. [5]

    Transunet: Rethinking the u-net architec- ture design for medical image segmentation through the lens of transformers.Medical Image Analysis, 97:103280, 2024

    Jieneng Chen, Jieru Mei, Xianhang Li, Yongyi Lu, Qihang Yu, Qingyue Wei, Xiangde Luo, Yutong Xie, Ehsan Adeli, Yan Wang, et al. Transunet: Rethinking the u-net architec- ture design for medical image segmentation through the lens of transformers.Medical Image Analysis, 97:103280, 2024. 12

    work page 2024

  6. [6]

    Simple baselines for image restoration.arXiv preprint arXiv:2204.04676, pages 17–33, 2022

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration.arXiv preprint arXiv:2204.04676, pages 17–33, 2022. 6, 10

    arXiv 2022

  7. [7]

    Computational optics for mobile terminals in mass production.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4245–4259, 2022

    Shiqi Chen, Ting Lin, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Computational optics for mobile terminals in mass production.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4245–4259, 2022. 6

    work page 2022

  8. [8]

    Dynamic convolu- tion: Attention over convolution kernels

    Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolu- tion: Attention over convolution kernels. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11030–11039, 2020. 6

    work page 2020

Show all 68 references
  1. [9]

    NTIRE 2025 challenge on image super-resolution (×4): Methods and results

    Zheng Chen, Kai Liu, Jue Gong, Jingkai Wang, Lei Sun, Zongwei Wu, Radu Timofte, Yulun Zhang, et al. NTIRE 2025 challenge on image super-resolution (×4): Methods and results. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- shops, 2025. 2

    2025

  2. [10]

    NTIRE 2025 challenge on real-world face restoration: Methods and results

    Zheng Chen, Jingkai Wang, Kai Liu, Jue Gong, Lei Sun, Zongwei Wu, Radu Timofte, Yulun Zhang, et al. NTIRE 2025 challenge on real-world face restoration: Methods and results. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- shops, 2025. 2

    2025

  3. [11]

    NTIRE 2025 challenge on raw image restoration and super-resolution

    Marcos Conde, Radu Timofte, et al. NTIRE 2025 challenge on raw image restoration and super-resolution. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025. 2

    2025

  4. [12]

    Raw image reconstruc- tion from RGB on smartphones

    Marcos Conde, Radu Timofte, et al. Raw image reconstruc- tion from RGB on smartphones. NTIRE 2025 challenge re- port. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) Workshops,

    2025

  5. [13]

    Model-based image signal processors via learnable dictionaries

    Marcos V Conde, Steven McDonagh, Matteo Maggioni, Ales Leonardis, and Eduardo P ´erez-Pellitero. Model-based image signal processors via learnable dictionaries. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 481–489, 2022. 1

    2022

  6. [14]

    Reversed image sig- nal processing and raw reconstruction

    Marcos V Conde, Radu Timofte, Yibin Huang, Jingyang Peng, Chang Chen, Cheng Li, Eduardo P´erez-Pellitero, Fen- glong Song, Furui Bai, Shuai Liu, et al. Reversed image sig- nal processing and raw reconstruction. aim 2022 challenge report. InEuropean Conference on Computer Visio...

    2022

  7. [15]

    Bsraw: Improving blind raw image super-resolution

    Marcos V Conde, Florin Vasluianu, and Radu Timofte. Bsraw: Improving blind raw image super-resolution. InPro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 8500–8510, 2024. 1

    2024

  8. [16]

    To- ward efficient deep blind raw image restoration

    Marcos V Conde, Florin Vasluianu, and Radu Timofte. To- ward efficient deep blind raw image restoration. In2024 IEEE International Conference on Image Processing (ICIP), pages 1725–1731. IEEE, 2024. 1

    2024

  9. [17]

    Vision transformers need registers.arXiv preprint arXiv:2309.16588, 2023

    Timoth ´ee Darcet, Maxime Oquab, Julien Mairal, and Pi- otr Bojanowski. Vision transformers need registers.arXiv preprint arXiv:2309.16588, 2023. 6

    2023 arXiv

  10. [18]

    Mobile computational photography: A tour.arXiv preprint arXiv:2102.09000, 2021

    Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Mobile computational photography: A tour.arXiv preprint arXiv:2102.09000, 2021. 1, 2

    2021 arXiv

  11. [19]

    1m parameters are enough? a lightweight cnn-based model for medical image segmentation

    Binh-Duong Dinh, Thanh-Thu Nguyen, Thi-Thao Tran, and Van-Truong Pham. 1m parameters are enough? a lightweight cnn-based model for medical image segmentation. In2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 127...

    2023

  12. [20]

    Rispnet: a network for reversed image signal pro- cessing

    Xiaoyi Dong, Yu Zhu, Chenghua Li, Peisong Wang, and Jian Cheng. Rispnet: a network for reversed image signal pro- cessing. InEuropean Conference on Computer Vision, pages 445–457. Springer, 2022. 1, 4

    2022

  13. [21]

    NTIRE 2025 challenge on night photography rendering

    Egor Ershov, Sergey Korchagin, Alexei Khalin, Artyom Pan- shin, Arseniy Terekhin, Ekaterina Zaychenkova, Georgiy Lobarev, Vsevolod Plokhotnyuk, Denis Abramov, Elisey Zhdanov, Sofia Dorogova, Yasin Mamedov, Nikola Banic, Georgii Perevozchikov, Radu Timofte, et al. NTIRE 2025 ch...

    2025

  14. [22]

    NTIRE 2025 challenge on cross-domain few-shot object detection: Methods and results

    Yuqian Fu, Xingyu Qiu, Bin Ren Yanwei Fu, Radu Timofte, Nicu Sebe, Ming-Hsuan Yang, Luc Van Gool, et al. NTIRE 2025 challenge on cross-domain few-shot object detection: Methods and results. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (...

    2025

  15. [23]

    Deep joint demosaicking and denoising.ACM Transactions on Graphics (ToG), 35(6):1–12, 2016

    Micha ¨el Gharbi, Gaurav Chaurasia, Sylvain Paris, and Fr´edo Durand. Deep joint demosaicking and denoising.ACM Transactions on Graphics (ToG), 35(6):1–12, 2016. 1

    2016

  16. [24]

    NTIRE 2025 challenge on text to image generation model quality assess- ment

    Shuhao Han, Haotian Fan, Fangyuan Kong, Wenjie Liao, Chunle Guo, Chongyi Li, Radu Timofte, et al. NTIRE 2025 challenge on text to image generation model quality assess- ment. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) Workshops,

    2025

  17. [25]

    Replac- ing mobile camera isp with a single deep learning model

    Andrey Ignatov, Luc Van Gool, and Radu Timofte. Replac- ing mobile camera isp with a single deep learning model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 536–537,

  18. [26]

    Learned smartphone isp on mobile gpus with deep learning, mobile ai & aim 2022 challenge: report

    Andrey Ignatov, Radu Timofte, Shuai Liu, Chaoyu Feng, Furui Bai, Xiaotao Wang, Lei Lei, Ziyao Yi, Yan Xiang, Zibin Liu, et al. Learned smartphone isp on mobile gpus with deep learning, mobile ai & aim 2022 challenge: report. InEuropean Conference on Computer Vision, pages 44–7...

    2022

  19. [27]

    NTIRE 2025 challenge on video quality enhancement for video con- ferencing: Datasets, methods and results

    Varun Jain, Zongwei Wu, Quan Zou, Louis Florentin, Hen- rik Turbell, Sandeep Siddhartha, Radu Timofte, et al. NTIRE 2025 challenge on video quality enhancement for video con- ferencing: Datasets, methods and results. InProceedings of the IEEE/CVF Conference on Computer Vision ...

    2025

  20. [28]

    Perceptual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer, 2016. 11

    2016

  21. [29]

    A software platform for manipulating the camera imaging pipeline

    Hakki Can Karaimer and Michael S Brown. A software platform for manipulating the camera imaging pipeline. In ECCV, pages 429–444, 2016. 1

    2016

  22. [30]

    NTIRE 2025 challenge on efficient burst hdr and restoration: Datasets, methods, and results

    Sangmin Lee, Eunpil Park, Angel Canelo, Hyunhee Park, Youngjo Kim, Hyungju Chun, Xin Jin, Chongyi Li, Chun-Le Guo, Radu Timofte, et al. NTIRE 2025 challenge on efficient burst hdr and restoration: Datasets, methods, and results. In Proceedings of the IEEE/CVF Conference on Com...

    2025

  23. [31]

    Omni-dimensional dynamic convolution.arXiv preprint arXiv:2209.07947,

    Chao Li, Aojun Zhou, and Anbang Yao. Omni-dimensional dynamic convolution.arXiv preprint arXiv:2209.07947,

    arXiv

  24. [32]

    NTIRE 2025 challenge on day and night raindrop removal for dual-focused images: Methods and results

    Xin Li, Yeying Jin, Xin Jin, Zongwei Wu, Bingchen Li, Yufei Wang, Wenhan Yang, Yu Li, Zhibo Chen, Bihan Wen, Robby Tan, Radu Timofte, et al. NTIRE 2025 challenge on day and night raindrop removal for dual-focused images: Methods and results. InProceedings of the IEEE/CVF Confe...

    2025

  25. [33]

    NTIRE 2025 challenge on short-form ugc video quality assessment and enhancement: Kwaisr dataset and study

    Xin Li, Xijun Wang, Bingchen Li, Kun Yuan, Yizhen Shao, Suhang Yao, Ming Sun, Chao Zhou, Radu Timofte, and Zhibo Chen. NTIRE 2025 challenge on short-form ugc video quality assessment and enhancement: Kwaisr dataset and study. InProceedings of the IEEE/CVF Conference on Com- pu...

    2025

  26. [34]

    NTIRE 2025 challenge on short-form ugc video quality assessment and enhancement: Methods and results

    Xin Li, Kun Yuan, Bingchen Li, Fengbin Guan, Yizhen Shao, Zihao Yu, Xijun Wang, Yiting Lu, Wei Luo, Suhang Yao, Ming Sun, Chao Zhou, Zhibo Chen, Radu Timofte, et al. NTIRE 2025 challenge on short-form ugc video quality assessment and enhancement: Methods and results. InPro- ce...

    2025

  27. [35]

    NTIRE 2025 the 2nd restore any image model (RAIM) in the wild challenge

    Jie Liang, Radu Timofte, Qiaosi Yi, Zhengqiang Zhang, Shuaizheng Liu, Lingchen Sun, Rongyuan Wu, Xindong Zhang, Hui Zeng, Lei Zhang, et al. NTIRE 2025 the 2nd restore any image model (RAIM) in the wild challenge. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion a...

    2025

  28. [36]

    Joint demo- saicing and denoising with self guidance

    Lin Liu, Xu Jia, Jianzhuang Liu, and Qi Tian. Joint demo- saicing and denoising with self guidance. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 4, 6

    2020

  29. [37]

    NTIRE 2025 XGC quality assessment chal- lenge: Methods and results

    Xiaohong Liu, Xiongkuo Min, Qiang Hu, Xiaoyun Zhang, Jie Guo, et al. NTIRE 2025 XGC quality assessment chal- lenge: Methods and results. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025. 2

    2025

  30. [38]

    NTIRE 2025 challenge on low light image enhancement: Methods and results

    Xiaoning Liu, Zongwei Wu, Florin-Alexandru Vasluianu, Hailong Yan, Bin Ren, Yulun Zhang, Shuhang Gu, Le Zhang, Ce Zhu, Radu Timofte, et al. NTIRE 2025 challenge on low light image enhancement: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion ...

    2025

  31. [39]

    Image semantic segmentation approach based on deeplabv3 plus network with an attention mech- anism.Engineering Applications of Artificial Intelligence, 127:107260, 2024

    Yanyan Liu, Xiaotian Bai, Jiafei Wang, Guoning Li, Jin Li, and Zengming Lv. Image semantic segmentation approach based on deeplabv3 plus network with an attention mech- anism.Engineering Applications of Artificial Intelligence, 127:107260, 2024. 12

    2024

  32. [40]

    Sgdr: Stochas- tic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2016

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochas- tic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2016. 8

    2016 arXiv

  33. [41]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 4, 8

    2017 arXiv

  34. [42]

    Rang M. H. Nguyen and Michael S. Brown. Raw image reconstruction using a self-contained srgb-jpeg image with only 64 kb overhead. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1655–1663,

  35. [43]

    Spatially aware metadata for raw reconstruction

    Abhijith Punnappurath and Michael S Brown. Spatially aware metadata for raw reconstruction. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 218–226, 2021. 1

    2021

  36. [44]

    Re- thinking the pipeline of demosaicing, denoising and super- resolution.arXiv preprint arXiv:1905.02538, 2019

    Guocheng Qian, Yuanhao Wang, Chao Dong, Jimmy S Ren, Wolfgang Heidrich, Bernard Ghanem, and Jinjin Gu. Re- thinking the pipeline of demosaicing, denoising and super- resolution.arXiv preprint arXiv:1905.02538, 2019. 1

    1905 arXiv

  37. [45]

    The tenth NTIRE 2025 efficient super- resolution challenge report

    Bin Ren, Hang Guo, Lei Sun, Zongwei Wu, Radu Timo- fte, Yawei Li, et al. The tenth NTIRE 2025 efficient super- resolution challenge report. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025. 2

    2025

  38. [46]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 6

    2015 arXiv

  39. [47]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMICCAI, pages 234–241. Springer, 2015. 3, 8, 9

    2015

  40. [48]

    NTIRE 2025 challenge on UGC video enhancement: Meth- ods and results

    Nickolay Safonov, Alexey Bryntsev, Andrey Moskalenko, Dmitry Kulikov, Dmitriy Vatolin, Radu Timofte, et al. NTIRE 2025 challenge on UGC video enhancement: Meth- ods and results. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- sh...

    2025

  41. [49]

    Deepisp: Toward learning an end-to-end image processing pipeline

    Eli Schwartz, Raja Giryes, and Alex M Bronstein. Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing, 28(2):912–923,

  42. [50]

    Donghwan Seo, Abhijith Punnappurath, Luxi Zhao, Abdel- rahman Abdelhamed, Sai Kiran Tedla, Sanguk Park, Jihwan Choe, and Michael S. Brown. Graphics2raw: Mapping com- puter graphics images to sensor raw images. InProceedings of the IEEE/CVF International Conference on Computer ...

    2023

  43. [51]

    Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network

    Wenzhe Shi, Jose Caballero, Ferenc Husz ´ar, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on compu...

    2016

  44. [52]

    Cyclical learning rates for training neural networks

    Leslie N Smith. Cyclical learning rates for training neural networks. In2017 IEEE winter conference on applications of computer vision (WACV), pages 464–472. IEEE, 2017. 4

    2017

  45. [53]

    NTIRE 2025 challenge on event-based image deblurring: Methods and results

    Lei Sun, Andrea Alfarano, Peiqi Duan, Shaolin Su, Kaiwei Wang, Boxin Shi, Radu Timofte, Danda Pani Paudel, Luc Van Gool, et al. NTIRE 2025 challenge on event-based image deblurring: Methods and results. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2025

  46. [54]

    The tenth ntire 2025 image denoising challenge report

    Lei Sun, Hang Guo, Bin Ren, Luc Van Gool, Radu Timo- fte, Yawei Li, et al. The tenth ntire 2025 image denoising challenge report. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025. 2

    2025

  47. [55]

    NTIRE 2025 image shadow removal challenge report

    Florin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou, Cailian Chen, Zongwei Wu, Radu Timofte, et al. NTIRE 2025 image shadow removal challenge report. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025. 2

    2025

  48. [56]

    NTIRE 2025 ambi- ent lighting normalization challenge

    Florin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou, Zongwei Wu, Radu Timofte, et al. NTIRE 2025 ambi- ent lighting normalization challenge. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025. 2

    2025

  49. [57]

    NTIRE 2025 challenge on light field image super-resolution: Methods and results

    Yingqian Wang, Zhengyu Liang, Fengyuan Zhang, Lvli Tian, Longguang Wang, Juncheng Li, Jungang Yang, Radu Timofte, Yulan Guo, et al. NTIRE 2025 challenge on light field image super-resolution: Methods and results. InPro- ceedings of the IEEE/CVF Conference on Computer Vision an...

    2025

  50. [58]

    Uformer: A general u-shaped transformer for image restoration

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683–17693, 2022. 4

    2022

  51. [59]

    Invertible im- age signal processing

    Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible im- age signal processing. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 6287–6296, 2021. 1

    2021

  52. [60]

    Condconv: Conditionally parameterized convolu- tions for efficient inference.Advances in neural information processing systems, 32, 2019

    Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolu- tions for efficient inference.Advances in neural information processing systems, 32, 2019. 6

    2019

  53. [61]

    NTIRE 2025 challenge on single image reflection removal in the wild: Datasets, methods and results

    Kangning Yang, Jie Cai, Ling Ouyang, Florin-Alexandru Vasluianu, Radu Timofte, Jiaming Ding, Huiming Sun, Lan Fu, Jinlong Li, Chiu Man Ho, Zibo Meng, et al. NTIRE 2025 challenge on single image reflection removal in the wild: Datasets, methods and results. InProceedings of the...

    2025

  54. [62]

    NTIRE 2025 challenge on hr depth from images of specular and transparent surfaces

    Pierluigi Zama Ramirez, Fabio Tosi, Luigi Di Stefano, Radu Timofte, Alex Costanzino, Matteo Poggi, Samuele Salti, Ste- fano Mattoccia, et al. NTIRE 2025 challenge on hr depth from images of specular and transparent surfaces. InPro- ceedings of the IEEE/CVF Conference on Comput...

    2025

  55. [63]

    Cycleisp: Real image restoration via improved data synthesis, 2020

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Cycleisp: Real image restoration via improved data synthesis, 2020. 1

    2020

  56. [64]

    Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary

    Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856– 286...

    2024

  57. [65]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InCVPR, pages 586–595,

  58. [66]

    Kbnet: Kernel basis network for image restoration.arXiv preprint arXiv:2303.02881, 2023

    Yi Zhang, Dasong Li, Xiaoyu Shi, Dailan He, Kangning Song, Xiaogang Wang, Honwei Qin, and Hongsheng Li. Kbnet: Kernel basis network for image restoration.arXiv preprint arXiv:2303.02881, 2023. 6

    2023 arXiv

  59. [67]

    Eednet: enhanced encoder-decoder network for autoisp

    Yu Zhu, Zhenyu Guo, Tian Liang, Xiangyu He, Chenghua Li, Cong Leng, Bo Jiang, Yifan Zhang, and Jian Cheng. Eednet: enhanced encoder-decoder network for autoisp. In Computer Vision–ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 171–

    2020

  60. [68]

    Rawhdr: High dynamic range image reconstruction from a single raw im- age

    Yunhao Zou, Chenggang Yan, and Ying Fu. Rawhdr: High dynamic range image reconstruction from a single raw im- age. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 12334–12344, 2023. 3, 4

    2023

Pith tools

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