pith. sign in

arxiv: 2605.22455 · v1 · pith:UNOOVLHZnew · submitted 2026-05-21 · 💻 cs.CV · cs.AI· cs.LG· physics.optics

Making the Discrete Continuous: Synthetic RAW Augmentations for Fine-Grained Evaluation of Person Detection Performance in Low Light

Valeria Pais , Malena Mendilaharzu , Daniele Faccio , Luis Oala , Christoph Clausen , Bruno Sanguinetti This is my paper

Pith reviewed 2026-05-22 06:53 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LGphysics.optics
keywords synthetic RAW augmentationlow-light detectionperson detectionnoise modelobject detectionautonomous drivingperformance evaluation
0
0 comments X

The pith

Synthetic RAW augmentations generate low-light images whose detection performance matches real low-light captures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how synthetic low-light samples generated from RAW augmentations can better characterize the performance of object detection models as a function of scene illumination. Real datasets are sparse in low-light conditions, limiting detailed evaluations for applications like autonomous driving. The synthetic technique matches the camera sensor noise model, producing performance metrics similar to real low-light data. This similarity means the AI model struggles to tell synthetic and real low-light inputs apart, supporting more continuous sampling of the input space for benchmarks.

Core claim

Focusing on pedestrian detection in the dark, the work uses a synthetic RAW image augmentation technique to generate low-light samples that match the noise model of the camera sensor. Performance metrics on real and synthetic low-light data are similar, indicating that the AI model finds it hard to distinguish between them.

What carries the argument

Synthetic RAW image augmentation technique to generate low-light samples that match the noise model of the camera sensor.

If this is right

  • Object detection performance can be evaluated continuously across different illumination levels using generated samples.
  • Synthetic data can improve coverage in low-density regions of real datasets for more reliable benchmarks.
  • The approach allows fine-grained analysis of model behavior in safety-critical low-light scenarios without additional real data collection.
  • Similar performance metrics validate the use of synthetics for testing in autonomous driving contexts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar techniques could be applied to other sensors or environmental conditions to create continuous evaluation spaces.
  • Accurate noise modeling might allow synthetic data to replace some real captures in model training pipelines.
  • A direct test would involve varying the augmentation parameters and measuring if detection accuracy remains aligned with real data.

Load-bearing premise

The synthetic RAW augmentation accurately reproduces the camera sensor's noise model across varying illumination levels without introducing artifacts that would affect detection differently than real low-light capture.

What would settle it

A controlled experiment comparing detection performance on paired real and synthetic images at the same illumination level and content would show divergence if the claim does not hold.

Figures

Figures reproduced from arXiv: 2605.22455 by Bruno Sanguinetti, Christoph Clausen, Daniele Faccio, Luis Oala, Malena Mendilaharzu, Valeria Pais.

Figure 1
Figure 1. Figure 1: (a) AODRaw sliced test dataset distribution of person instances as a function of bounding box mean electrons and area. (b) Real [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Person detection performance on real data as a function [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Person detection performance on synthetic data as a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Experiment to determine whether the detector can dis [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Person detection performance on real data as a function [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Person detection performance on synthetic data with [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Person detection performance on real and synthetic low [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
read the original abstract

Real-world deployment of AI vision models is both fueled and limited by the data available for training and testing. Real datasets are sparse and uneven: long-tailed or unbalanced distributions hinder generalization, and the low number of samples in low density regions makes it hard to run evaluations. Synthetic data can fill these gaps, providing us with a way to sample the input space more continuously and improve data coverage for benchmarks. Focusing on the autonomous driving safety-critical case of pedestrian detection in the dark, we show how synthetic low-light samples can be used to better characterize the performance of a state-of-the-art object detection model as a function of the scene illumination. We use a synthetic RAW image augmentation technique to generate low-light samples that match the noise model of the camera sensor. Performance metrics on real and synthetic low-light data are similar, indicating that the AI model finds it hard to distinguish between them.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces a synthetic RAW image augmentation technique that generates low-light samples matching a camera sensor's noise model. Focused on pedestrian detection in dark conditions for autonomous driving, it compares a state-of-the-art detector's performance metrics (e.g., mAP, recall) on real low-light images versus the synthetically augmented ones. The central claim is that similar metrics across real and synthetic data indicate the model finds it hard to distinguish the domains, enabling finer-grained, continuous evaluation of detection performance as a function of scene illumination.

Significance. If the core results hold after addressing the interpretation gap, the work provides a practical method to densely sample illumination levels where real data is sparse, improving benchmark granularity for safety-critical vision tasks. The physics-based noise modeling is a strength when properly validated, as it avoids purely data-driven artifacts and supports reproducible augmentations.

major comments (2)
  1. [Abstract and §5] Abstract and §5 (Discussion/Conclusion): The inference that similar performance metrics on real vs. synthetic low-light data indicate the detector 'finds it hard to distinguish between them' is not supported by metric parity alone. Comparable mAP or recall can result from independent failure modes (e.g., noise statistics or artifacts degrading bounding-box regression equally) without interchangeable internal representations or decision boundaries. The paper should add direct evidence such as feature-space distances, domain classification accuracy on detector embeddings, or adversarial domain tests to substantiate this load-bearing claim for the evaluation-utility argument.
  2. [§3] §3 (Methods): The noise model includes free scaling parameters; the manuscript must specify their selection or fitting procedure (e.g., via sensor measurements or cross-validation on held-out real pairs) and report quantitative fidelity metrics (noise variance, histogram matching) across illumination levels to rule out post-hoc tuning that could artifactually align detection outcomes.
minor comments (2)
  1. [Results tables] Table 2 or equivalent results table: Include per-illumination bin breakdowns with confidence intervals to clarify the fine-grained evaluation claim.
  2. [Figures] Figure 3 (example images): Add side-by-side quantitative comparisons (e.g., PSNR or noise power spectra) between real and synthetic patches for visual clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. The comments highlight important aspects of our claims and methodological transparency. We address each major comment below, proposing targeted revisions where appropriate to strengthen the manuscript without altering its core contributions.

read point-by-point responses

  1. Referee: [Abstract and §5] Abstract and §5 (Discussion/Conclusion): The inference that similar performance metrics on real vs. synthetic low-light data indicate the detector 'finds it hard to distinguish between them' is not supported by metric parity alone. Comparable mAP or recall can result from independent failure modes (e.g., noise statistics or artifacts degrading bounding-box regression equally) without interchangeable internal representations or decision boundaries. The paper should add direct evidence such as feature-space distances, domain classification accuracy on detector embeddings, or adversarial domain tests to substantiate this load-bearing claim for the evaluation-utility argument.

    Authors: We agree that metric similarity alone does not rigorously demonstrate that the detector's internal representations or decision boundaries are interchangeable between domains. Our central contribution is the utility of synthetic data for densely sampling illumination levels where real data is sparse, and the observed parity in mAP and recall across real and synthetic samples supports that the augmentations are sufficiently realistic for this evaluation purpose. We will revise the abstract and §5 to remove the phrasing that the model 'finds it hard to distinguish between them' and instead emphasize that performance parity indicates the synthetic data can serve as a proxy for fine-grained analysis. We will also add a short discussion of limitations regarding direct domain-invariance evidence and note that future work could include embedding-based analyses. revision: partial

  2. Referee: [§3] §3 (Methods): The noise model includes free scaling parameters; the manuscript must specify their selection or fitting procedure (e.g., via sensor measurements or cross-validation on held-out real pairs) and report quantitative fidelity metrics (noise variance, histogram matching) across illumination levels to rule out post-hoc tuning that could artifactually align detection outcomes.

    Authors: We acknowledge that the current description of the noise model parameters is insufficiently detailed. In the revised manuscript, we will explicitly document the procedure for selecting and fitting the free scaling parameters, including whether they were derived from sensor specifications, calibrated on real image pairs, or validated via cross-validation. We will also add quantitative fidelity results, such as comparisons of noise variance and intensity histogram matching between real and synthetic images at multiple illumination levels, to demonstrate that the model was not tuned post-hoc to match detection outcomes. revision: yes

Circularity Check

0 steps flagged

No significant circularity; evaluation uses independent held-out real data

full rationale

The paper describes generating synthetic low-light RAW images via an augmentation technique that matches the camera sensor noise model drawn from physics. It then reports that detection performance metrics on these synthetics are similar to those on real low-light images, from which it infers the model has difficulty distinguishing domains. This comparison is to held-out real data rather than data used to define or fit the augmentation itself. No equations or steps reduce a claimed prediction or first-principles result to the inputs by construction, and no self-citation chain or ansatz smuggling is evident in the provided derivation. The similarity observation is an empirical outcome, not a tautological re-statement of the generation process.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the chosen noise model faithfully represents real sensor behavior at low light and that the detection model treats synthetic and real images equivalently. No free parameters are explicitly named in the abstract, but the augmentation process likely involves at least one scaling factor for photon noise or read noise that must be calibrated to the target camera.

free parameters (1)
  • noise model scaling parameters
    Parameters controlling photon shot noise and read noise levels in the synthetic augmentation must be set to match a specific sensor; these are fitted or chosen to align synthetic and real distributions.
axioms (1)
  • domain assumption The camera sensor noise can be modeled as a combination of signal-dependent and signal-independent components that remain valid under synthetic illumination reduction.
    Invoked when generating low-light samples that are claimed to match real capture.

pith-pipeline@v0.9.0 · 5710 in / 1441 out tokens · 28513 ms · 2026-05-22T06:53:05.091633+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 1 internal anchor

  1. [1]

    Dual-conditioned temporal diffusion modeling for driving scene generation

    Xiangyu Bai, Yedi Luo, Le Jiang, and Sarah Ostadabbas. Dual-conditioned temporal diffusion modeling for driving scene generation. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 8412–8419, 2025. 2

    work page 2025

  2. [2]

    ReRAW: RGB-to-RAW image recon- struction 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. InProceedings of the Computer Vision and Pat- tern Recognition Conference (CVPR), pages 11833–11843,

    work page

  3. [3]

    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. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3

    work page 2019

  4. [4]

    Cascade R-CNN: Delv- ing into high quality object detection

    Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN: Delv- ing into high quality object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2018. 3

    work page 2018

  5. [5]

    MMDetection: Open MMLab Detection Toolbox and Benchmark

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tian- heng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and...

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  6. [6]

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

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

    work page 2023

  7. [7]

    End-to-end autonomous driving: Challenges and frontiers.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10164– 10183, 2024

    Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, An- dreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10164– 10183, 2024. 2

    work page 2024

  8. [8]

    Conde, Steven McDonagh, Matteo Maggioni, Ales Leonardis, and Eduardo P ´erez-Pellitero

    Marcos V . Conde, Steven McDonagh, Matteo Maggioni, Ales Leonardis, and Eduardo P ´erez-Pellitero. Model-based image signal processors via learnable dictionaries.Proceed- ings of the AAAI Conference on Artificial Intelligence, 36(1): 481–489, 2022. 2

    work page 2022

  9. [9]

    David F. Crouse. On implementing 2d rectangular assign- ment algorithms.IEEE Transactions on Aerospace and Elec- tronic Systems, 52(4):1679–1696, 2016. 2

    work page 2016

  10. [10]

    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. InComputer Vision – ECCV 2024, pages 37–56, Cham, 2025. Springer Nature Switzerland. 2, 3

    work page 2024

  11. [11]

    Multitask aet with orthogo- nal tangent regularity for dark object detection

    Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogo- nal tangent regularity for dark object detection. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2533–2542, 2021. 2

    work page 2021

  12. [12]

    Brown, and Peyman Milanfar

    Mauricio Delbracio, Damien Kelly, Michael S. Brown, and Peyman Milanfar. Mobile computational photography: A tour.Annual Review of Vision Science, 7(V olume 7, 2021): 571–604, 2021. 2

    work page 2021

  13. [13]

    ImageNet: A large-scale hierarchical im- age database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical im- age database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2

    work page 2009

  14. [14]

    CARLA: An open urban driving simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. InProceedings of the 1st Annual Conference on Robot Learning, pages 1–16. PMLR, 2017. 2

    work page 2017

  15. [15]

    Hasinoff, and William T

    Fredo Durand, Samuel W. Hasinoff, and William T. Free- man. Noise-optimal capture for high dynamic range photog- raphy. In2010 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 553–560, Los Alamitos, CA, USA, 2010. IEEE Computer Society. 3 5

    work page 2010

  16. [16]

    MBLLEN: Low-light image/video enhancement using cnns

    Jianhua Wu Feifan Lv, Feng Lu and Chongsoon Lim. MBLLEN: Low-light image/video enhancement using cnns. British Machine Vision Conference, 2018. 2

    work page 2018

  17. [17]

    Practical poissonian-gaussian noise mod- eling and fitting for single-image raw-data.IEEE Transac- tions on Image Processing, 17(10):1737–1754, 2008

    Alessandro Foi, Mejdi Trimeche, Vladimir Katkovnik, and Karen Egiazarian. Practical poissonian-gaussian noise mod- eling and fitting for single-image raw-data.IEEE Transac- tions on Image Processing, 17(10):1737–1754, 2008. 3

    work page 2008

  18. [18]

    Craft- ing object detection in very low light.Proceedings of the British Machine Vision Conference 2021, 2021

    Yang Hong, Kaixuan Wei, Linwei Chen, and Ying Fu. Craft- ing object detection in very low light.Proceedings of the British Machine Vision Conference 2021, 2021. 2

    work page 2021

  19. [19]

    Kong Li, Zhe Dai, Xuan Wang, Yongchao Song, and Gwang- gil Jeon. GAN-Based controllable image data augmentation in low-visibility conditions for improved roadside traffic per- ception.IEEE Transactions on Consumer Electronics, 70(3): 6174–6188, 2024. 2

    work page 2024

  20. [20]

    Salman Asif, and Zhan Ma

    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

    work page 2024

  21. [21]

    Towards RAW object detection in di- verse conditions

    Zhong-Yu Li, Xin Jin, Bo-Yuan Sun, Chun-Le Guo, and Ming-Ming Cheng. Towards RAW object detection in di- verse conditions. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 8859–8868,

    work page

  22. [22]

    Lawrence Zitnick

    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. InComputer Vision – ECCV 2014, pages 740–755, Cham,

    work page 2014

  23. [23]

    Springer International Publishing. 3

    work page

  24. [24]

    Learning raw image denoising with bayer pattern unification and bayer preserving augmen- tation

    Jiaming Liu, Chi-Hao Wu, Yuzhi Wang, Qin Xu, Yuqian Zhou, Haibin Huang, Chuan Wang, Shaofan Cai, Yifan Ding, Haoqiang Fan, and Jue Wang. Learning raw image denoising with bayer pattern unification and bayer preserving augmen- tation. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2070– 2077, 2019. 2

    work page 2070

  25. [25]

    Practical signal-dependent noise parameter estimation from a single noisy image.IEEE Transactions on Image Process- ing, 23(10):4361–4371, 2014

    Xinhao Liu, Masayuki Tanaka, and Masatoshi Okutomi. Practical signal-dependent noise parameter estimation from a single noisy image.IEEE Transactions on Image Process- ing, 23(10):4361–4371, 2014. 3

    work page 2014

  26. [26]

    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. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 11976– 11986, 2022. 3

    work page 2022

  27. [27]

    Text2Traffic: a text-to-image generation and editing method for traffic scenes

    Feng Lv, Haoxuan Feng, Zilu Zhang, Chunlong Xia, and Yanfeng Li. Text2Traffic: a text-to-image generation and editing method for traffic scenes. InInternational Confer- ence on AI-Generated Content (AIGC 2025), page 141830B. International Society for Optics and Photonics, SPIE, 2026. 2

    work page 2025

  28. [28]

    Data models for dataset drift controls in machine learning with optical images.Transactions on Machine Learning Re- search, 2023

    Luis Oala, Marco Aversa, Gabriel Nobis, Kurt Willis, Yoan Neuenschwander, Mich `ele Buck, Christian Matek, Jerome Extermann, Enrico Pomarico, Wojciech Samek, Roderick Murray-Smith, Christoph Clausen, and Bruno Sanguinetti. Data models for dataset drift controls in machine learning with optical images.Transactions on Machine Learning Re- search, 2023. 2

    work page 2023

  29. [29]

    PAS- CALRAW: Raw image database for object detection

    Alex Omid-Zohoor, David Ta, and Boris Murmann. PAS- CALRAW: Raw image database for object detection. Stan- ford Digital Repository, 2015. Accessed: 2026-04-14. 2

    work page 2015

  30. [30]

    Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St ´efan J

    Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St ´efan J. van der Walt, Matthew Brett, Joshua Wil- son, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, ˙Ilhan Polat, Yu Feng, Eric ...

    work page 2020

  31. [31]

    Invertible image signal processing

    Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 6287–6296, 2021. 2

    work page 2021

  32. [32]

    UniSim: A neural closed-loop sensor simulator

    Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Mani- vasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Ur- tasun. UniSim: A neural closed-loop sensor simulator. In 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 1389–1399, 2023. 2

    work page 2023

  33. [33]

    Rawgment: Noise-Accounted RAW Aug- mentation Enables Recognition in a Wide Variety of Envi- ronments

    Masakazu Yoshimura, Junji Otsuka, Atsushi Irie, and Takeshi Ohashi. Rawgment: Noise-Accounted RAW Aug- mentation Enables Recognition in a Wide Variety of Envi- ronments . In2023 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 14007–14017, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 2

    work page 2023

  34. [34]

    CycleISP: Real image restoration via improved data synthesis

    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. In2020 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 2693–2702,

    work page

  35. [35]

    ChatScene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles

    Jiawei Zhang, Chejian Xu, and Bo Li. ChatScene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15459–15469, 2024. 2 6 Making the Discrete Continuous: Synthetic RA W Augmentations for Fine-Grained Evaluation of Person Detection ...

    work page 2024