Pith. sign in

REVIEW 3 major objections 5 minor 58 references

FreqAdapt: Frequency-Adaptive Processing for RAW Object Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read FreqAdapt processes RAW camera images in the Fourier domain, separating ISP operations into amplitude and phase branches, and reports the best object-detection results across five RAW benchmarks while adding only +0.019M parameters.

desk verdict The empirical story mostly holds, but the 'physics-inspired' framing is mathematically wrong: gamma and CCM do not act as amplitude-only/phase-only in Fourier space, so the interpretability claim collapses while the learned module may still work as a heuristic enhancer. read the letter →

arxiv 2608.03385 v1 pith:45E7YOMO submitted 2026-08-04 cs.CV

Hanxi Li , Huiling Li This is my paper
classification cs.CV
keywords RAWobjectdetectionfrequencydomainimagesignalprocessor(ISP)amplitude-phasedecompositionadaptiveenhancementlow-lightplug-and-playmoduleFouriertransform
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 proposes FreqAdapt, a lightweight add-on module that enhances RAW sensor images for object detection by processing them in the Fourier frequency domain. Its central claim is that ISP operations split naturally into two families: intensity adjustments (white balance, brightness, gamma, noise reduction) act on the amplitude spectrum, while structural and color adjustments (color correction, sharpening, detail enhancement) act on the phase spectrum. A small frequency encoder reads amplitude, phase, and RAW statistics to predict the parameters of both branches, and a learned fusion combines the enhanced image with the original before passing it to any standard detector. On the LOD, NOD, and AROD benchmarks, FreqAdapt reports state-of-the-art mAP with only +0.019M parameters and +2.25 GFLOPs, and it also improves segmentation when plugged into a pretrained detector. The paper's motivation is that RAW data preserves sensor information that sRGB pipelines discard, and that frequency-domain processing can recover it more cheaply than spatial-domain learned ISPs.

What carries the argument

The central object is the 2D discrete Fourier transform of the RAW image, decomposed as $F(I)=A\cdot e^{i\Phi}$, with the amplitude $A$ and phase $\Phi$ processed by separate ISP branches. The operative mechanism is amplitude-phase domain separation: intensity-related ISP operations are implemented as multiplicative or thresholding masks on $A$, while structure-related operations are implemented as phase shifts and additions on $\Phi$, followed by inverse FFT and learned adaptive fusion with the original RAW image. The Fourier decomposition gives the module a global receptive field and turns expensive spatial operations into element-wise spectral operations.

What would settle it

Take a RAW image $I$, apply a true gamma curve $\gamma(\cdot)$ in the spatial domain, and compare $F(\gamma(I))$ with the paper's amplitude-only model $|F(I)|^{\gamma}\cdot e^{i\angle F(I)}$; if the phases or cross-frequency amplitudes differ beyond numerical precision, Eq. (5) is not the Fourier dual of gamma correction. Likewise, apply a 3×3 CCM in the spatial domain and check whether the amplitude spectrum changes; the paper's phase-only CCM of Eq. (9) predicts that it should not.

Watch

Extended reading notes

Core claim

FreqAdapt's core discovery is that a Fourier decomposition of RAW images—writing the spectrum as $F(I)=A\cdot e^{i\Phi}$—lets the ISP be decoupled into an amplitude branch and a phase branch, so each classic ISP step is executed as a simple spectral modulation rather than a sequential spatial pipeline. The amplitude branch multiplies and thresholds spectral magnitudes to implement white balance, brightness, gamma, and denoising; the phase branch shifts spectral phases to implement color correction, sharpening, and detail enhancement. A Frequency RAW Encoder pools amplitude, phase, and RAW image statistics into a global conditioning vector that predicts all ISP parameters per image, and an ad

Load-bearing premise

The load-bearing premise is that gamma correction and color correction can each be enacted in the Fourier domain as pure amplitude and pure phase operations respectively, even though spatial gamma is a pointwise nonlinearity and spatial color correction mixes channels, so both mathematically perturb both spectra.

Editorial extensions

If this is right

  • On the five benchmark datasets used (LOD-Dark, LOD-Normal, NOD-Nikon, NOD-Sony, AROD), FreqAdapt reports higher mAP than the previous best method RAM by 0.2–1.1 mAP, with the largest gain on NOD-Sony.
  • The module adds only +0.019M parameters and +2.25 GFLOPs, so a detector can gain roughly 2.3 mAP over raw input at a fraction of the cost of scaling the backbone from ResNet18 to ResNet50.
  • FreqAdapt also improves detection when the downstream detector is frozen (pretrained), indicating that it can adapt RAW inputs to sRGB-trained features.
  • The same module transfers to semantic segmentation on the ADE20K RAW dataset, improving mIoU under dark, normal, and over-exposed conditions.
  • RAW inputs benefit more than sRGB inputs from FreqAdapt (e.g., 29.4 vs 27.7 mAP on NOD-Nikon), consistent with the claim that unprocessed sensor data retains more useful frequency information.

Reading between the lines

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

  • The clean amplitude/phase story is an approximation: gamma correction is a pointwise nonlinearity whose Fourier transform is a convolution, so it genuinely changes both amplitude and phase, and color correction mixes channels in a way that also perturbs both spectra. The empirical gains may not depend on the physics story being literally exact, so a useful test would compare FreqAdapt against a ve
  • Because the reported margins over RAM are small (0.2–1.1 mAP), the paper's single-run results would be more convincing with repeated-seed variance and significance testing; this is not reported.
  • The module's frequency-selective denoising and phase sharpening suggest a natural extension to RAW burst or HDR inputs, where combining spectral enhancement with physically calibrated noise models might compound the gains; the paper does not test this.
  • The paper itself notes limited spatial flexibility for non-uniform lighting, so a direct stress test would apply FreqAdapt to scenes with strong local illumination gradients, where global frequency masks may underperform spatially adaptive ISPs.
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. FreqAdapt proposes a lightweight, plug-and-play module for RAW object detection that operates in the Fourier domain. The input RAW image is transformed via FFT, then processed by an amplitude branch (white balance, brightness, gamma, noise reduction) and a phase branch (color correction matrix, sharpening, detail enhancement), augmented by an adaptive frequency encoder and a learnable fusion step. The paper claims that this amplitude/phase separation is based on the physical properties of ISP operations and therefore has good physical interpretability. Experiments on LOD, NOD, and AROD report state-of-the-art mAP with small parameter and FLOP overhead, plus ablations and a semantic-segmentation transfer study.

Significance. If the empirical results hold, FreqAdapt is a lightweight module with consistent gains over prior methods such as RAM across multiple RAW detection benchmarks, and the efficiency profile is attractive. The paper also includes component ablations and a spatial-domain comparison that support the empirical design. However, the central conceptual claim—that gamma correction and CCM are naturally phase-preserving or phase-only in the Fourier domain—is mathematically incorrect. Gamma is a pointwise nonlinearity and changes both amplitude and phase; CCM is linear channel mixing that mixes complex spectra. The stated 'physics-inspired domain separation' therefore lacks a sound basis, although the module could still work as a learned frequency-domain enhancement. The inconsistencies between Tables 1 and 2 further weaken the presentation. With the theoretical claims reframed and the experimental reporting clarified, the empirical contribution is potentially publishable.

major comments (3)
  1. [Section 1 and Eq. (5)] The claim that gamma correction is an amplitude-domain, phase-preserving operation is not supported. Gamma is a pointwise nonlinearity; its Fourier transform is a convolution in the frequency domain, so it changes both amplitude and phase. Applying A_bright^gamma to the amplitude spectrum in Eq. (5) is not equivalent to spatial gamma correction and artificially preserves phase. This is a load-bearing premise for the amplitude/phase domain separation, so the stated physical interpretability collapses unless a correct derivation or an explicit heuristic justification is supplied.
  2. [Section 3.2.3, Eq. (9), and Section C] Spatial CCM is a per-pixel linear channel mixing. In the Fourier domain, each output channel is F(output_i)=sum_j c_ij F(input_j), which mixes complex coefficients and therefore changes both amplitude and phase. Equation (9), which reshapes and applies a 3x3 matrix to the phase vector, is not the Fourier-domain realization of any color correction matrix. The cross-power spectrum argument in Section C only shows that relative phase encodes inter-channel correlation; it does not establish that CCM acts only on phase. The phase-branch design and the 'physics-inspired' domain separation therefore lack a mathematical basis.
  3. [Tables 1 and 2] The same method is reported with different scores on NOD-Sony: FreqAdapt 27.3/53.7 in Table 1 vs 27.6/55.2 in Table 2, and FreqAdapt-T 26.6/50.9 vs 26.7/54.1. If this is due to different detectors (Faster R-CNN/ResNet50 vs RetinaNet/ResNet18), the text should say so explicitly where the numbers are cited. As written, Section 4.3.2 cites Table 2's 27.6 mAP as FreqAdapt's NOD-Sony performance, which can be mistaken for Table 1's result and affects the state-of-the-art and efficiency claims.
minor comments (5)
  1. [Table 1] The GenISP NOD-Sony mAP50 entry contains a comma ('51,8') instead of a decimal point. Also, RAW-Adapter is listed with '-' on AROD; please state explicitly that this baseline was not evaluated on AROD.
  2. [Eq. (4)] Multiplying the amplitude spectrum by the Gaussian mask M_low corresponds to a convolution in the spatial domain, not a simple global brightness scaling. Please clarify whether this is intended as brightness enhancement or low-pass filtering.
  3. [Eq. (10)] Applying a Laplacian to wrapped phase values is ill-defined near the +/-pi boundary. Please describe how phase wrapping is handled before computing ∇^2(Φ_ccm), or whether the gradient is computed in the complex domain.
  4. [Appendix E] The text says FreqAdapt-T removes noise reduction, sharpening, and detail enhancement, but Table 9 still lists Phase Adjuster and Edge Enhancer for FreqAdapt-T. Clarify whether these are simplified variants or should be marked as inactive.
  5. [Supplementary Section A] The dataset name 'AODRAW' appears where AROD is intended. Please check for similar typos throughout the supplementary material.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical performance comparisons and a learned enhancement module, neither of which reduces to its inputs by construction.

full rationale

I find no circular derivation. FreqAdapt's central claim is the empirical mAP comparison against baselines on LOD, NOD, and AROD (Sec. 4.3.1, Tables 1-2), which is self-contained: the module's parameters are learned end-to-end from detection loss, not fitted to the reported benchmark numbers, and no equation in the paper forces the observed gains. The design premise that gamma correction and white balance are amplitude-domain operations and CCM is a phase-domain operation (Sec. 1, Sec. C, Eqs. 5 and 9) is a modeling assumption, not a circular argument; even if the Fourier-domain characterization is mathematically questionable (gamma is a pointwise nonlinearity that alters both amplitude and phase, and CCM mixes complex channel coefficients), that is a correctness/validity risk, not a reduction of the output to the input. There are no load-bearing self-citations: the reference list contains no prior work by the present authors (Hanxi Li, Huiling Li), and no uniqueness theorem or ansatz is imported from the authors' own previous papers. The supplementary limitations (Sec. D) explicitly acknowledge training instability, spatial inflexibility, and deployment constraints, and these do not constitute circularity. The reported performance improvements are externally falsifiable against the specified baselines and datasets, so the paper is not circular; the score is 0.

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

The central empirical claim relies on the learnable network parameters (fitted to data), but the conceptual claim about domain separation rests on the questionable axioms above; the gamma and CCM axioms are not supported and in fact are contradicted by Fourier properties.

free parameters (2)
  • Gaussian mask bandwidth alpha
    Hand-chosen hyperparameter in Eq. 4 defining the low-frequency mask M_low; value not reported.
  • Frequency band masks (M_mid, M_high, M_detail)
    Hand-designed filters in Eqs. 7, 10, 11; cutoffs not specified.
assumptions (4)
  • standard math Fourier decomposition F(I) = A exp(i Phi)
    Standard Fourier transform property used in Eq. 2.
  • domain assumption Noise predominantly occupies high-frequency amplitude
    Used in Section 3.1 to justify amplitude-domain noise reduction; plausible but not tested.
  • ad hoc to paper Gamma correction preserves phase and acts multiplicatively on amplitude
    This is false for pointwise nonlinear transforms; asserted without proof in Eq. 5.
  • ad hoc to paper CCM corresponds to phase differences between channels
    Cross-power spectrum phase is not equivalent to CCM action; asserted without rigorous derivation in Section C, Eq. 15.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FreqAdapt: Frequency-Adaptive Processing for RAW Object Detection." pith.science (2026). https://pith.science/paper/45E7YOMO

@misc{pith2026260803385,
  author       = {Pith},
  title        = {Pith review of: FreqAdapt: Frequency-Adaptive Processing for RAW Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/45E7YOMO}},
  note         = {Machine review of arXiv:2608.03385}
}
read the original abstract

Existing object detection methods predominantly utilize sRGB inputs, which are compressed from RAW sensor data using Image Signal Processors (ISP) originally designed for visualization purposes. Compared to RGB images, RAW images possess favorable noise characteristics and richer information representation, which are crucial for object detection, particularly under challenging conditions such as adverse weather or low-light environments. In this paper, we propose FreqAdapt, a lightweight module for adaptive RAW data enhancement in the frequency domain. Unlike traditional spatial domain processing methods, FreqAdapt innovatively maps ISP operations to the Fourier frequency domain and performs domain separation based on the physical properties of ISP operations, ensuring each operation is performed in its most suitable domain. Meanwhile, through an adaptive frequency domain encoder that jointly analyzes amplitude spectrum, phase spectrum, and RAW image features, we provide global context for ISP parameter prediction and employ a learnable fusion mechanism to achieve adaptive feature enhancement. Extensive experiments on multiple datasets with diverse lighting and weather conditions demonstrate that FreqAdapt achieves state-of-the-art performance while maintaining lightweight efficiency and good physical interpretability. Furthermore, our module can be seamlessly incorporated into existing object detection frameworks, providing a novel solution for visual perception tasks in the RAW domain.

Figures

Figures reproduced from arXiv: 2608.03385 by the authors.

Figure 1
Figure 1. (a) Illustration of the core framework of our proposed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture of FreqAdapt. Our framework decomposes RAW images into amplitude and phase spectra via FFT, processes them [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visual comparison of detection results across different methods on challenging scenarios from LOD-Dark (low-light), NOD-Sony [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on AROD dataset under three challenging weather conditions. From top to bottom: Low-Night, Fog, and [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Visualization of semantic segmentation results on ADE20K RAW under normal exposure conditions. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 51 canonical work pages

  1. [1]

    Prentice-Hall, Inc., 1988

    E Oran Brigham.The fast Fourier transform and its applica- tions. Prentice-Hall, Inc., 1988. 2

    work page 1988

  2. [2]

    Re- configuring the imaging pipeline for computer vision

    Mark Buckler, Suren Jayasuriya, and Adrian Sampson. Re- configuring the imaging pipeline for computer vision. In Proceedings of the IEEE International Conference on Com- puter Vision, pages 975–984, 2017. 3

    work page 2017

  3. [3]

    Cascade r-cnn: Delving into high quality object detection

    Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 6154–6162, 2018. 5

    work page 2018

  4. [4]

    Learn- ing to see in the dark

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learn- ing to see in the dark. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 3291–3300, 2018. 2

    work page 2018

  5. [5]

    Amplitude-phase recombination: Rethink- ing robustness of convolutional neural networks in frequency domain

    Guangyao Chen, Peixi Peng, Li Ma, Jia Li, Lin Du, and Yonghong Tian. Amplitude-phase recombination: Rethink- ing robustness of convolutional neural networks in frequency domain. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 458–467, 2021. 2

    work page 2021

  6. [6]

    Mmdetection: Open mmlab detection toolbox and benchmark.arXiv preprint arXiv:1906.07155, 2019

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark.arXiv preprint arXiv:1906.07155, 2019. 5

    arXiv 1906

  7. [7]

    Instance segmentation in the dark.International Journal of Computer Vision, 131(8):2198–2218, 2023

    Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark.International Journal of Computer Vision, 131(8):2198–2218, 2023. 2

    work page 2023

  8. [8]

    Raw-adapter: Adapting pre- trained visual model to camera raw images

    Ziteng Cui and Tatsuya Harada. Raw-adapter: Adapting pre- trained visual model to camera raw images. InEuropean Conference on Computer Vision, pages 37–56. Springer, 2025. 2, 3, 6

    work page 2025

Show all 58 references
  1. [9]

    Multitask aet with orthogonal tan- gent regularity for dark object detection

    Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tan- gent regularity for dark object detection. InProceedings of the IEEE/CVF international conference on computer vision, pages 2553–2562, 2021. 2

    2021

  2. [10]

    Awnet: Attentive wavelet network for image isp

    Linhui Dai, Xiaohong Liu, Chengqi Li, and Jun Chen. Awnet: Attentive wavelet network for image isp. InComputer Vision– ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 185–201. Springer, 2020. 2

    2020

  3. [11]

    Mobile computational photography: A tour.Annual review of vision science, 7(1):571–604, 2021

    Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Mobile computational photography: A tour.Annual review of vision science, 7(1):571–604, 2021. 2

    2021

  4. [12]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 1

    2009

  5. [13]

    Beyond rgb: Adaptive parallel processing for raw object detection.arXiv preprint arXiv:2503.13163,

    Shani Gamrian, Hila Barel, Feiran Li, Masakazu Yoshimura, and Daisuke Iso. Beyond rgb: Adaptive parallel processing for raw object detection.arXiv preprint arXiv:2503.13163,

    arXiv

  6. [14]

    Dark-isp: Enhancing raw image processing for low-light object detection

    Jiasheng Guo, Xin Gao, Yuxiang Yan, Guanghao Li, and Jian Pu. Dark-isp: Enhancing raw image processing for low-light object detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9583– 9593, 2025. 1, 2

    2025

  7. [15]

    Learn- ing degradation-independent representations for camera isp pipelines

    Yanhui Guo, Fangzhou Luo, and Xiaolin Wu. Learn- ing degradation-independent representations for camera isp pipelines. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25774– 25783, 2024. 1, 2

    2024

  8. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 5

    2016

  9. [17]

    Enhancing raw-to-srgb with decoupled style structure in fourier domain

    Xuanhua He, Tao Hu, Guoli Wang, Zejin Wang, Run Wang, Qian Zhang, Keyu Yan, Ziyi Chen, Rui Li, Chengjun Xie, et al. Enhancing raw-to-srgb with decoupled style structure in fourier domain. InProceedings of the AAAI Conference on Artificial Intelligence, pages 2130–2138, 2024. 2

    2024

  10. [18]

    Craft- ing object detection in very low light

    Yang Hong, Kaixuan Wei, Linwei Chen, and Ying Fu. Craft- ing object detection in very low light. InBMVC, page 3, 2021. 5

    2021

  11. [19]

    Aim 2019 challenge on raw to rgb mapping: Methods and results

    Andrey Ignatov, Radu Timofte, Sung-Jea Ko, Seung-Wook Kim, Kwang-Hyun Uhm, Seo-Won Ji, Sung-Jin Cho, Jun-Pyo Hong, Kangfu Mei, Juncheng Li, et al. Aim 2019 challenge on raw to rgb mapping: Methods and results. In2019 IEEE/CVF International Conference on Computer Vision Worksho...

    2019

  12. [20]

    Replacing mobile camera isp with a single deep learning model

    Andrey Ignatov, Luc Van Gool, and Radu Timofte. Replacing mobile camera isp with a single deep learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 536–537, 2020. 2

    2020

  13. [21]

    Lighting every darkness in two pairs: A calibration-free pipeline for raw denoising

    Xin Jin, Jia-Wen Xiao, Ling-Hao Han, Chunle Guo, Ruixun Zhang, Xialei Liu, and Chongyi Li. Lighting every darkness in two pairs: A calibration-free pipeline for raw denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13275–13284, 2023. 2

    2023

  14. [22]

    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 Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceed- ings, Part I 14, pages 429–444. Springer, 2016. 1, 2

    2016

  15. [23]

    D3net: Joint demosaick- ing, deblurring and deringing

    Tom´aˇs Kerepeck`y and Filip ˇSroubek. D3net: Joint demosaick- ing, deblurring and deringing. In2020 25th International Conference on Pattern Recognition (ICPR), pages 1–8. IEEE,

  16. [24]

    Paramisp: learned forward and inverse isps using camera parameters.arXiv preprint arXiv:2312.13313, 2023

    Woohyeok Kim, Geonu Kim, Junyong Lee, Seungyong Lee, Seung-Hwan Baek, and Sunghyun Cho. Paramisp: learned forward and inverse isps using camera parameters.arXiv preprint arXiv:2312.13313, 2023. 2

    2023 arXiv

  17. [25]

    Image features from phase congruency

    Peter Kovesi et al. Image features from phase congruency. Videre: Journal of computer vision research, 1(3):1–26, 1999. 3

    1999

  18. [26]

    High dynamic range and super-resolution from raw image bursts.arXiv preprint arXiv:2207.14671, 2022

    Bruno Lecouat, Thomas Eboli, Jean Ponce, and Julien Mairal. High dynamic range and super-resolution from raw image bursts.arXiv preprint arXiv:2207.14671, 2022. 2

    2022 arXiv

  19. [27]

    Efficient visual computing with camera raw snapshots.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 46(7):4684–4701, 2024

    Zhihao Li, Ming Lu, Xu Zhang, Xin Feng, M Salman Asif, and Zhan Ma. Efficient visual computing with camera raw snapshots.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 46(7):4684–4701, 2024. 2

    2024

  20. [28]

    Towards raw object detection in diverse conditions

    Zhong-Yu Li, Xin Jin, Bo-Yuan Sun, Chun-Le Guo, and Ming-Ming Cheng. Towards raw object detection in diverse conditions. InProceedings of the Computer Vision and Pat- tern Recognition Conference, pages 8859–8868, 2025. 2, 5

    2025

  21. [29]

    Raw image deblurring.IEEE Transactions on Multimedia, 24:61–72, 2020

    Chih-Hung Liang, Yu-An Chen, Yueh-Cheng Liu, and Win- ston H Hsu. Raw image deblurring.IEEE Transactions on Multimedia, 24:61–72, 2020. 2

    2020

  22. [30]

    Focal loss for dense object detection.arXiv preprint arXiv:1708.02002, 2017

    T Lin. Focal loss for dense object detection.arXiv preprint arXiv:1708.02002, 2017. 5

    2017 arXiv

  23. [31]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

    2014

  24. [32]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. ieee. InCVF Conference on Computer Vision and Pattern Recognition (CVPR), page 7, 2022. 5

    2022

  25. [33]

    Logarithmic lenses: Exploring log rgb data for image classification

    Bruce A Maxwell, Sumegha Singhania, Avnish Patel, Rahul Kumar, Heather Fryling, Sihan Li, Haonan Sun, Ping He, and Zewen Li. Logarithmic lenses: Exploring log rgb data for image classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2024

  26. [34]

    Genisp: Neural isp for low- light machine cognition

    Igor Morawski, Yu-An Chen, Yu-Sheng Lin, Shusil Dangi, Kai He, and Winston H Hsu. Genisp: Neural isp for low- light machine cognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 630–639, 2022. 2, 3, 5, 6, 7

    2022

  27. [35]

    Feature detection from local energy.Pattern recognition letters, 6(5):303–313,

    M Concetta Morrone and Robyn A Owens. Feature detection from local energy.Pattern recognition letters, 6(5):303–313,

  28. [36]

    Hardware-in-the- loop end-to-end optimization of camera image processing pipelines

    Ali Mosleh, Avinash Sharma, Emmanuel Onzon, Fahim Man- nan, Nicolas Robidoux, and Felix Heide. Hardware-in-the- loop end-to-end optimization of camera image processing pipelines. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7529–7538,

  29. [37]

    Automatic isp image quality tuning using nonlinear optimization

    Jun Nishimura, Timo Gerasimow, Rao Sushma, Aleksandar Sutic, Chyuan-Tyng Wu, and Gilad Michael. Automatic isp image quality tuning using nonlinear optimization. In2018 25th IEEE International Conference on Image Processing (ICIP), pages 2471–2475. IEEE, 2018. 2

    2018

  30. [38]

    The importance of phase in signals.Proceedings of the IEEE, 69(5):529–541, 2005

    Alan V Oppenheim and Jae S Lim. The importance of phase in signals.Proceedings of the IEEE, 69(5):529–541, 2005. 2

    2005

  31. [39]

    Learning to exploit the sequence-specific prior knowledge for image processing pipelines optimization

    Haina Qin, Longfei Han, Weihua Xiong, Juan Wang, Wen- tao Ma, Bing Li, and Weiming Hu. Learning to exploit the sequence-specific prior knowledge for image processing pipelines optimization. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, ...

    2023

  32. [40]

    Color image processing pipeline.IEEE Signal processing magazine, 22(1):34–43, 2005

    Rajeev Ramanath, Wesley E Snyder, Youngjun Yoo, and Mark S Drew. Color image processing pipeline.IEEE Signal processing magazine, 22(1):34–43, 2005. 1, 2

    2005

  33. [41]

    Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural information process- ing systems, 28, 2015

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural information process- ing systems, 28, 2015. 5

    2015

  34. [42]

    Raw image processing for python, a wrapper for libraw.Date accessed, 1(12):2022, 2014

    Maik Riechert. Raw image processing for python, a wrapper for libraw.Date accessed, 1(12):2022, 2014. 5

    2022

  35. [43]

    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,

  36. [44]

    Hyperparameter optimization in black-box im- age processing using differentiable proxies.ACM Trans

    Ethan Tseng, Felix Yu, Yuting Yang, Fahim Mannan, Karl ST Arnaud, Derek Nowrouzezahrai, Jean-Franc ¸ois Lalonde, and Felix Heide. Hyperparameter optimization in black-box im- age processing using differentiable proxies.ACM Trans. Graph., 38(4):27–1, 2019. 2

    2019

  37. [45]

    Adaptiveisp: Learning an adaptive image signal pro- cessor for object detection.Advances in Neural Information Processing Systems, 37:112598–112623, 2024

    Yujin Wang, Tianyi Xu, Zhang Fan, Tianfan Xue, and Jinwei Gu. Adaptiveisp: Learning an adaptive image signal pro- cessor for object detection.Advances in Neural Information Processing Systems, 37:112598–112623, 2024. 1, 2, 3

    2024

  38. [46]

    A physics-based noise formation model for extreme low-light raw denoising

    Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758– 2767, 2020. 2

    2020

  39. [47]

    Physics-based noise modeling for extreme low-light photog- raphy.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8520–8537, 2021

    Kaixuan Wei, Ying Fu, Yinqiang Zheng, and Jiaolong Yang. Physics-based noise modeling for extreme low-light photog- raphy.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8520–8537, 2021. 2

    2021

  40. [48]

    Visionisp: Repurposing the image signal processor for computer vision applications

    Chyuan-Tyng Wu, Leo F Isikdogan, Sushma Rao, Bhavin Nayak, Timo Gerasimow, Aleksandar Sutic, Liron Ain- Kedem, and Gilad Michael. Visionisp: Repurposing the image signal processor for computer vision applications. In 2019 IEEE International Conference on Image Processing (ICIP...

    2019

  41. [49]

    Simrod: A simple baseline for raw object detec- tion with global and local enhancements.arXiv preprint arXiv:2503.07101, 2025

    Haiyang Xie, Xi Shen, Shihua Huang, Qirui Wang, and Zheng Wang. Simrod: A simple baseline for raw object detec- tion with global and local enhancements.arXiv preprint arXiv:2503.07101, 2025. 1

    2025

  42. [50]

    Invertible image signal processing

    Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 6287–6296, 2021. 2

    2021

  43. [51]

    Learning in the frequency domain

    Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1740–1749, 2020. 2

    2020

  44. [52]

    Toward raw object detection: A new benchmark and a new model

    Ruikang Xu, Chang Chen, Jingyang Peng, Cheng Li, Yibin Huang, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Toward raw object detection: A new benchmark and a new model. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 13384–13393,

  45. [53]

    Dynamicisp: dynamically controlled image sig- nal processor for image recognition

    Masakazu Yoshimura, Junji Otsuka, Atsushi Irie, and Takeshi Ohashi. Dynamicisp: dynamically controlled image sig- nal processor for image recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12866–12876, 2023. 2, 3, 6

    2023

  46. [54]

    Reconfigisp: Reconfigurable camera image processing pipeline

    Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248–4257, 2021. 2, 3

    2021

  47. [55]

    Isp-teacher: Image signal process with disentanglement regularization for unsupervised domain adaptive dark object detection

    Yin Zhang, Yongqiang Zhang, Zian Zhang, Man Zhang, Rui Tian, and Mingli Ding. Isp-teacher: Image signal process with disentanglement regularization for unsupervised domain adaptive dark object detection. InProceedings of the AAAI Conference on Artificial Intelligence, pages 73...

    2024

  48. [56]

    Learning raw-to-srgb mappings with inaccurately aligned supervision

    Zhilu Zhang, Haolin Wang, Ming Liu, Ruohao Wang, Jiawei Zhang, and Wangmeng Zuo. Learning raw-to-srgb mappings with inaccurately aligned supervision. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4348–4358, 2021. 2

    2021

  49. [57]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Bar- riuso, and Antonio Torralba. Scene parsing through ade20k dataset. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 633–641, 2017. 2

    2017

  50. [58]

    His- togram of oriented gradients feature extraction from raw bayer pattern images.IEEE Transactions on Circuits and Systems II: Express Briefs, 67(5):946–950, 2020

    Wei Zhou, Shengyu Gao, Ling Zhang, and Xin Lou. His- togram of oriented gradients feature extraction from raw bayer pattern images.IEEE Transactions on Circuits and Systems II: Express Briefs, 67(5):946–950, 2020. 3 FreqAdapt: Frequency-Adaptive Processing for RA W Object Dete...

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.