REVIEW 4 major objections 3 minor 42 references
Spatial-Frequency Aware for Object Detection in RAW Image
T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read This paper claims that object detection on RAW images improves when frequency bands are converted back into spatial maps and fused with original spatial features through cross-domain attention, with per-domain adaptive gamma correction.
desk verdict An intriguing framework for RAW object detection, but the available record is just an abstract with no evidence, so it cannot be judged yet. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The key mechanism is frequency-band spatialization: the RAW image is decomposed into frequency bands, each band is inverse-transformed back into a spatial map, and those maps are fed alongside the original spatial features into a cross-domain attention module that fuses them. The same framework then predicts and applies separate gamma correction parameters for the spatial and frequency-derived domains, enabling adaptive per-domain nonlinear adjustment.
What would settle it
Replace the frequency-band spatial maps with fixed random maps of the same shape while keeping every other component identical; if detection performance does not drop relative to the full SFAE, then the claimed frequency information is not what drives the gain. A second check is to feed raw Fourier magnitude spectra into the same attention module instead of the spatialized bands; if the spatialized maps do not outperform the raw spectra, the spatialization step itself is not load-bearing.
Extended reading notes
Core claim
The central claim is that RAW image object detection is improved by spatializing frequency bands rather than manipulating abstract spectra. Each frequency band is inverse-transformed into a tangible spatial map, preserving physical intuition about contours and textures. A cross-domain fusion attention module then allows deep interactions between these maps and the original spatial features, and the framework performs adaptive nonlinear adjustments by predicting different gamma parameters for the two domains. The authors assert that this recovers object details suppressed by RAW's wide dynamic range and linear response, and that it outperforms methods that enhance only in the spatial domain.
Load-bearing premise
The load-bearing premise is that converting each frequency band back into a spatial map preserves the contours and textures that RAW's linear response suppresses, so the attention module can exploit them; if these maps carry nothing new, the entire frequency branch becomes redundant.
Editorial extensions
If this is right
- If the claim holds, RAW-image object detectors can gain accuracy without changing the detection backbone, simply by preprocessing with the SFAE enhancer.
- The cross-domain fusion attention design could be reused for other vision tasks that pair spatial maps with frequency-decomposed representations, such as dehazing or medical image enhancement.
- Per-domain gamma prediction implies that a single global enhancement curve is suboptimal for RAW images, and content-dependent, domain-specific curves become the standard approach.
- Spatializing frequency bands makes frequency-domain processing more interpretable, because the network operates on concrete edge and texture maps rather than abstract spectra.
Reading between the lines
- Editorial inference: the approach implicitly assumes that inverse-transformed frequency bands carry object-discriminative information not already present in the spatial branch; quantifying the redundancy between the two branches would isolate where the gain actually comes from.
- Editorial inference: if spatialization is the load-bearing step, the method may generalize to other linear-sensor modalities, such as hyperspectral or infrared RAW data, where wide dynamic range similarly suppresses detail.
- Editorial inference: the adaptive gamma prediction is label-free, so it could be tested as a standalone preprocessing module; comparing detector performance with and without the fusion attention would reveal how much of the benefit comes from the gamma adjustment alone.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript as available through arXiv:2508.01396 consists of an abstract describing SFAE, a framework for object detection on RAW images that spatializes frequency bands, fuses them with spatial features through cross-domain attention, and applies per-domain adaptive gamma correction. The abstract claims that this approach recovers object contours and textures suppressed by RAW's wide dynamic range and linear response. However, the available record contains no equations, no architecture details, no training protocol, no dataset description, and no experimental results. The supplied full text is a different paper on symbolic song generation (arXiv:2508.01394) and therefore does not provide any technical support for the claims made in this submission.
Significance. If the claimed method were validated, it would address a genuine limitation in RAW-based object detection: the linear response and wide dynamic range of RAW images can suppress fine object details, and frequency-domain enhancement with spatialization and cross-domain attention is a plausible way to recover such details. That said, the contribution is currently unsubstantiated. The paper as presented offers no quantitative or qualitative evidence, no ablation study, and no comparison against existing enhancement methods, so its practical significance cannot be assessed beyond the plausibility of the idea.
major comments (4)
- [Abstract, contribution bullets 1–3] The central claim that SFAE improves object detection on RAW images is not supported by the available record: the abstract contains no equations, no architecture details, no training protocol, no dataset description, and no experimental or ablation results. Consequently, there is no evidence that spatializing frequency bands, fusing them with spatial features via cross-domain attention, and applying per-domain gamma correction actually recovers the contours and textures suppressed by RAW's wide dynamic range and linear response. Controlled experiments and comparisons with existing enhancement methods are required before this claim can be credited.
- [Abstract, contribution bullet 1] The 'spatialization' of frequency bands is the load-bearing assumption: the abstract asserts that inverse-transforming individual frequency bands into spatial maps preserves direct physical intuition and enables the attention module to exploit discriminative contours and textures, but it does not demonstrate that this process retains the localization and phase information needed for object detection. Without a derivation, an implementation description, or a sanity-check experiment, the motivation for the architecture is unverified.
- [Abstract, contribution bullet 3] The abstract does not specify how the per-domain gamma parameters are supervised or whether they are predicted independently of the detection loss. If they are learned end-to-end by optimizing detection performance, the term 'predicting' is misleading, and the contribution reduces to a learned nonlinearity; if they are obtained from a separate mechanism, that mechanism is absent. In either case, the claimed advantage over a single learned enhancement is unsupported by any comparison.
- [Full text (supplied)] The full text supplied with this review is not the SFAE manuscript but a separate paper on bar-level symbolic song generation (arXiv:2508.01394). As a result, the equations, architecture diagrams, experimental setup, and results that would normally be used to evaluate the central claim are unavailable. This mismatch must be corrected before a substantive review can occur.
minor comments (3)
- [Title and abstract] The title uses 'Spatial-Frequency Aware' while the abstract and contribution text use 'Space-Frequency Aware'; the terminology should be made consistent.
- [Abstract] The abstract would benefit from at least one quantitative result or a clear statement that this is a framework/position paper; as written, the reader cannot distinguish a completed system from a proposal.
- [Abstract] No references are given to prior RAW-enhancement or frequency-domain detection methods, which makes it difficult to assess the claimed novelty of SFAE.
Circularity Check
No significant circularity in the available abstract; the claims are unverified but not self-referential or fitted.
full rationale
The target manuscript (arXiv:2508.01396) is represented in the available record only by its abstract; the supplied full text belongs to a different submission (arXiv:2508.01394, a song-generation paper), so no equations, architecture details, training protocol, or experimental comparisons for SFAE can be inspected. From the abstract alone, the three contributions—inverse-transforming frequency bands into spatial maps, cross-domain fusion attention, and per-domain adaptive gamma—are architectural proposals rather than derivations. There is no fitted parameter that is later renamed as a prediction, no self-citation invoked as load-bearing evidence, and no stated quantity defined in terms of another claimed output. The assertion that frequency-domain decomposition recovers contours and textures suppressed by RAW's dynamic range is an empirical hypothesis that needs controlled ablation, but it is not circular: it does not reduce to the definition of any input or output. The mismatched full text is a documentation defect, not circularity. Therefore the circularity score is 0.
Assumptions & free parameters
assumptions (3)
- domain assumption RAW images have a wide dynamic range and linear response that suppress object details.
- domain assumption Object contours and textures can be naturally separated based on frequency.
- domain assumption Inversely transforming individual frequency bands into spatial maps preserves useful physical intuition and is trainable or discriminative.
Cite this review
Pith. "Pith review of Spatial-Frequency Aware for Object Detection in RAW Image." pith.science (2026). https://pith.science/paper/JPMHPLW7
@misc{pith2026250801396,
author = {Pith},
title = {Pith review of: Spatial-Frequency Aware for Object Detection in RAW Image},
year = {2026},
howpublished = {\url{https://pith.science/paper/JPMHPLW7}},
note = {Machine review of arXiv:2508.01396}
}
read the original abstract
Direct RAW-based object detection offers great promise by utilizing RAW data (unprocessed sensor data), but faces inherent challenges due to its wide dynamic range and linear response, which tends to suppress crucial object details. In particular, existing enhancement methods are almost all performed in the spatial domain, making it difficult to effectively recover these suppressed details from the skewed pixel distribution of RAW images. To address this limitation, we turn to the frequency domain, where features, such as object contours and textures, can be naturally separated based on frequency. In this paper, we propose Space-Frequency Aware RAW Image Object Detection Enhancer (SFAE), a novel framework that synergizes spatial and frequency representations. Our contribution is threefold. The first lies in the ``spatialization" of frequency bands. Different from the traditional paradigm of directly manipulating abstract spectra in deep networks, our method inversely transforms individual frequency bands back into tangible spatial maps, thus preserving direct physical intuition. Then the cross-domain fusion attention module is developed to enable deep multimodal interactions between these maps and the original spatial features. Finally, the framework performs adaptive nonlinear adjustments by predicting and applying different gamma parameters for the two domains.
Reference graph
Works this paper leans on
-
[1]
11, A. S. 2004. Specification for octave-band and fractional-octave-band analog and digital filters
work page 2004
-
[2]
Afifi, M.; Abdelhamed, A.; Abuolaim, A.; Punnappurath, A.; and Brown, M. S. 2021. Cie xyz net: Unprocessing images for low-level computer vision tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9): 4688--4700
work page 2021
-
[3]
Brooks, T.; Mildenhall, B.; Xue, T.; Chen, J.; Sharlet, D.; and Barron, J. T. 2019. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11036--11045
work page 2019
-
[4]
Buckler, M.; Jayasuriya, S.; and Sampson, A. 2017. Reconfiguring the imaging pipeline for computer vision. In Proceedings of the IEEE International Conference on Computer Vision, 975--984
work page 2017
-
[5]
Chen, C.; Chen, Q.; Xu, J.; and Koltun, V. 2018. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, 3291--3300
2018
-
[6]
Chen, H.; and Ma, K. 2022. LW-ISP: A Lightweight Model with ISP and Deep Learning. arXiv preprint arXiv:2210.03904
work page Pith review arXiv 2022
-
[7]
Chen, K.; Wang, J.; Pang, J.; Cao, Y.; Xiong, Y.; Li, X.; Sun, S.; Feng, W.; Liu, Z.; Xu, J.; et al. 2019. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155
arXiv 2019
-
[8]
V.; McDonagh, S.; Maggioni, M.; Leonardis, A.; and P \'e rez-Pellitero, E
Conde, M. V.; McDonagh, S.; Maggioni, M.; Leonardis, A.; and P \'e rez-Pellitero, E. 2022. Model-based image signal processors via learnable dictionaries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 481--489
work page 2022
Show all 42 references
-
[9]
Cui, Z.; and Harada, T. 2024. RAW-adapter: Adapting pre-trained visual model to camera RAW images. In European Conference on Computer Vision, 37--56. Springer
2024
-
[10]
Dai, L.; Liu, X.; Li, C.; and Chen, J. 2020. Awnet: Attentive wavelet network for image isp. In European Conference on Computer Vision, 185--201. Springer
2020
-
[11]
Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, 248--255. Ieee
2009
-
[12]
He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. 770--778
2016
-
[13]
He, X.; Hu, T.; Wang, G.; Wang, Z.; Wang, R.; Zhang, Q.; Yan, K.; Chen, Z.; Li, R.; Xie, C.; et al. 2024. Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2130--2138
2024
-
[14]
Hong, Y.; Wei, K.; Chen, L.; and Fu, Y. 2021. Crafting Object Detection in Very Low Light
2021
-
[15]
Ignatov, A.; Chiang, C.-M.; Kuo, H.-K.; Sycheva, A.; and Timofte, R. 2021. Learned smartphone isp on mobile npus with deep learning, mobile ai 2021 challenge: Report. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2503--2514
2021
-
[16]
Ignatov, A.; Timofte, R.; Liu, S.; Feng, C.; Bai, F.; Wang, X.; Lei, L.; Yi, Z.; Xiang, Y.; Liu, Z.; et al. 2022. Learned smartphone ISP on mobile GPUs with deep learning, mobile AI & AIM 2022 challenge: report. In European Conference on Computer Vision, 44--70. Springer
2022
-
[17]
Ignatov, A.; Van Gool, L.; and Timofte, R. 2020. Replacing mobile camera isp with a single deep learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 536--537
2020
-
[18]
Jocher, G.; Chaurasia, A.; and Qiu, J. 2023. Ultralytics YOLOv8
2023
-
[19]
Klambauer, G.; Unterthiner, T.; Mayr, A.; and Hochreiter, S. 2017. Self-normalizing neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, 972–981. Red Hook, NY, USA: Curran Associates Inc. ISBN 9781510860964
2017
-
[20]
Lamba, M.; and Mitra, K. 2021. Restoring Extremely Dark Images in Real Time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 3487--3497
2021
-
[21]
Li, C.; Guo, C.-L.; Zhou, M.; Liang, Z.; Zhou, S.; Feng, R.; and Loy, C. C. 2023. Embedding Fourier for Ultra-High-Definition Low-Light Image Enhancement. In ICLR
2023
-
[22]
S.; and Ma, Z
Li, Z.; Lu, M.; Zhang, X.; Feng, X.; Asif, M. S.; and Ma, Z. 2024. Efficient visual computing with camera raw snapshots. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7): 4684--4701
2024
-
[23]
Lin, T.-Y.; Goyal, P.; Girshick, R.; He, K.; and Doll \'a r, P. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, 2980--2988
2017
-
[24]
Liu, W.; Ren, G.; Yu, R.; Guo, S.; Zhu, J.; and Zhang, L. 2022. Image-adaptive YOLO for object detection in adverse weather conditions. In Proceedings of the AAAI conference on artificial intelligence, volume 36, 1792--1800
2022
-
[25]
Ljungbergh, W.; Johnander, J.; Petersson, C.; and Felsberg, M. 2023. Raw or cooked? object detection on raw images. In Scandinavian Conference on Image Analysis, 374--385. Springer
2023
-
[26]
Lu, Q.; and Murmann, B. 2024. Enhancing the energy efficiency and robustness of TinyML computer vision using coarsely-quantized log-gradient input images. ACM Transactions on Embedded Computing Systems, 23(3): 1--20
2024
-
[27]
Morawski, I.; Chen, Y.-A.; Lin, Y.-S.; Dangi, S.; He, K.; and Hsu, W. H. 2022. Genisp: Neural isp for low-light machine cognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 630--639
2022
-
[28]
Omid-Zohoor, A.; Young, C.; Ta, D.; and Murmann, B. 2017. Toward always-on mobile object detection: Energy versus performance tradeoffs for embedded HOG feature extraction. IEEE Transactions on Circuits and Systems for Video Technology, 28(5): 1102--1115
2017
-
[29]
Oppenheim, A.; and Schafer, R. 1975. Digital Signal Processing. Prentice Hall international editions. Prentice-Hall. ISBN 9789380618388
1975
-
[30]
W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al
Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR
2021
-
[31]
Redmon, J.; Divvala, S.; Girshick, R.; and Farhadi, A. 2016. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 779--788
2016
-
[32]
Ren, S.; He, K.; Girshick, R.; and Sun, J. 2016. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6): 1137--1149
2016
-
[33]
Sun, P.; Zhang, R.; Jiang, Y.; Kong, T.; Xu, C.; Zhan, W.; Tomizuka, M.; Li, L.; Yuan, Z.; Wang, C.; et al. 2021. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14454--14463
2021
-
[34]
Suvorov, R.; Logacheva, E.; Mashikhin, A.; Remizova, A.; Ashukha, A.; Silvestrov, A.; Kong, N.; Goka, H.; Park, K.; and Lempitsky, V. 2022. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of co...
2022
-
[35]
Ulyanov, D.; Vedaldi, A.; and Lempitsky, V. 2016. Instance Normalization: The Missing Ingredient for Fast Stylization. arXiv preprint arXiv:1607.08022
2016 arXiv
-
[36]
N.; Kaiser, .; and Polosukhin, I
Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. volume 30
2017
-
[37]
Woo, S.; Park, J.; Lee, J.-Y.; and Kweon, I. S. 2018. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), 3--19
2018
-
[38]
Xu, R.; Chen, C.; Peng, J.; Li, C.; Huang, Y.; Song, F.; Yan, Y.; and Xiong, Z. 2023. Toward raw object detection: A new benchmark and a new model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 13384--13393
2023
-
[39]
W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F
Zamir, S. W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F. S.; Yang, M.-H.; and Shao, L. 2020. Cycleisp: Real image restoration via improved data synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2696--2705
2020
-
[40]
Zhou, M.; Huang, J.; Guo, C.-L.; and Li, C. 2023. Fourmer: An efficient global modeling paradigm for image restoration. In International conference on machine learning, 42589--42601. PMLR
2023
-
[41]
, " * write output.state after.block = add.period write newline
ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...
-
[42]
write newline
" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.