Pith. sign in

REVIEW 4 major objections 5 minor 36 references

RAISE: Realness Assessment for Image Synthesis and Evaluation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper introduces RAISE, a 600-image dataset with human realness scores, and shows that pretrained vision features can predict those scores.

desk verdict Useful small dataset, but the modeling claims are confounded by real-vs-generated separability and a sloppy ablation table. read the letter →

arxiv 2505.19233 v2 pith:4LTHMA6Q submitted 2025-05-25 cs.CV cs.AIcs.MMeess.IV

classification cs.CVcs.AIcs.MMeess.IV
keywords AI-generatedimagesperceptualrealnesssubjectivequalityassessmenthumanstudydatasettransferlearningfoundationvisionmodelsmeanopinionscore
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 introduces RAISE, a dataset of 600 images (480 AI-generated and 120 real photographs) with human ratings of perceived realness collected from 47 participants in a controlled study. The central claim is that features extracted from deep foundation vision models pretrained on large image collections can predict these subjective realness judgments: a small regressor on frozen ResNet-18 features reaches a Spearman correlation of 0.6684 with the human scores on the test set. This matters because it suggests an image can be scored for realness without needing its text prompt or any task-specific training, which existing quality and naturalness models require. The paper also shows that real images are rated more real on average than AI-generated ones, although the two groups overlap substantially.

What carries the argument

The key machinery is the pairing of subjective mean opinion scores (MOS) obtained from a single-stimulus human study with transfer-learned deep features. The MOS values, averaged over at least 23 raters per image after outlier removal, provide a continuous ground-truth scale of realness from 0 to 100. The learned features come from a pretrained ResNet-18 backbone whose global-average-pooled 512-dimensional output is fed into a four-layer fully connected regressor, with the backbone kept fixed; this compact setup is what demonstrates that realness information is already present in generic foundation-model representations.

What would settle it

Re-evaluate all models under repeated random train/test splits or leave-one-out cross-validation on RAISE, and check whether the ranking of models and the reported SROCC values remain stable; alternatively, test the best model on an independent set of freshly generated images and see if its SROCC against human realness ratings stays well above zero.

Watch

Extended reading notes

Core claim

The discovery is that perceptual realness, a subjective quality distinct from technical image quality, is captured to a useful degree by generic features from deep vision backbones. Using the RAISE dataset's mean opinion scores as ground truth, the paper compares five approaches: handcrafted color, edge, contrast, texture, and frequency features with a decision tree (SROCC 0.5033); a lightweight CNN trained from scratch (0.6028); a regressor on frozen ImageNet-pretrained ResNet-18 features (0.6684); the existing JOINT naturalness model used as is (0.2998); and the JOINT rationality branch with partial fine-tuning (0.6798). The steady improvement from handcrafted features to foundation-model features is the evidence for the paper's claim that deep foundation vision models can effectively capture subjective realness.

Load-bearing premise

The comparisons between models rely on a single division of the 600 images into 510 training and 90 test images, so the small differences in correlation (for instance 0.6798 versus 0.6684) could vanish or swap under a different split.

Editorial extensions

If this is right

  • Realness scores can be assigned to any single image without access to the text prompt that generated it, unlike prompt-dependent quality metrics.
  • A small regressor on frozen pretrained features outperforms both handcrafted features and a CNN trained from scratch on 600 images, suggesting foundation-model representations are a strong default for perceptual tasks with limited data.
  • The near-zero transfer of JOINT to RAISE indicates that technical naturalness and perceived realness are not the same thing, and that realness needs its own benchmark.
  • The dataset provides a training and evaluation resource for realness models, and the authors propose using such models to guide generative models toward fewer unrealistic artifacts.

Reading between the lines

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

  • Because real and AI-generated images overlap in rated realness, a model trained on RAISE is probably not a reliable forgery detector; it ranks plausibility, not provenance.
  • The same kind of frozen-feature regressor could be adapted to rank or filter diffusion outputs during sampling, offering a parameter-free way to steer generation toward higher perceived realness.
  • The reported differences between the top models rely on one fixed split, so the specific ordering of the transfer-learning variants should be treated as provisional until tested across multiple splits or an independent set.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces RAISE, a dataset of 600 images (480 AI-generated and 120 real photographs) annotated with subjective perceptual-realness scores from a 47-subject human study, and reports baseline models for predicting those scores. The baselines include a decision-tree regressor on handcrafted color, edge, contrast, texture/structure, and frequency features; a lightweight CNN trained from scratch; a neural regressor on frozen ResNet-18 features; and transfer-learning variants of the JOINT model. The central claim is that features from deep foundation vision models can effectively capture subjective realness.

Significance. If the central claim is properly established, RAISE would be a useful community resource: it focuses on perceptual realness rather than technical quality, it is prompt-agnostic, and it includes real images as reference anchors. The subjective study follows ITU recommendation procedures, reports inter-subject consistency, and releases the data and models publicly. However, the current evidence for the central claim is weakened by an uncontrolled source-label confound, a single small test split without significance assessment, and an internal inconsistency in the reported decision-tree results. These issues are fixable but load-bearing, because the paper's headline conclusion is about capturing subjective realness, not about separating real from generated images.

major comments (4)
  1. [Section V-A, Tables II and IV] There is an internal inconsistency in the reported decision-tree performance. Table II reports SROCC = 0.5033 for the model using all 55 handcrafted features, but Table IV reports SROCC = 0.5106 for the same decision-tree model. Moreover, the text states that the ablation study shows 'consistent drop in performance' when any feature category is excluded, yet Table II shows that excluding the Color category improves SROCC from 0.5033 to 0.5106. This directly contradicts the claimed consistency and needs to be corrected, not merely re-worded, because the ablation is used to support the contribution of each feature group.
  2. [Section V, Tables III and IV] All conclusions about model ranking rely on a single partition into 510 training and 90 test images, with no confidence intervals, bootstrap estimates, cross-validation, or significance tests. For example, the reported difference between the ResNet-18 transfer model (SROCC 0.6684) and the fine-tuned JOINT rationality model (SROCC 0.6798) may be within sampling noise on a 90-image test set. The paper should report uncertainty around the SROCC/PLCC values (e.g., bootstrap intervals or repeated random splits) before claiming that one model 'conclusively outperformed' another.
  3. [Section IV and Section V-D] The mixed real/AIGI test set confounds subjective realness with source-label separability. The test partition mixes AIGIs and real images, where real images have a mean MOS of 76.56 and AIGIs average 60.98, with a large FID of 177.07 between the subsets. A model that mostly learns to separate real images from AIGIs can therefore achieve a high SROCC on the mixed test set without capturing fine-grained realness within either category. To support the abstract's claim, the paper needs at least one of the following: (i) report SROCC/PLCC separately on the real-only and AIGI-only test subsets; (ii) compare against a trivial binary source classifier as a baseline; or (iii) report partial correlations controlling for the real/generated source label. Without such a control, the reported correlations (e.g., 0.6798) do not establish that the models assess perceived realness rather than synthetic-vs-real category.
  4. [Section V-C and V-D] The comparison between the frozen ResNet-18 feature extractor and the fine-tuned JOINT rationality branch is not matched in a way that isolates the contribution of the foundation-model features. The two models use different backbones (ResNet-18 vs ResNet-50), different numbers of trainable parameters, and different fine-tuning protocols. Since the paper's central contribution is the dataset rather than a new metric, this is not fatal, but the conclusion that 'features derived from deep foundation vision models can effectively capture the subjective realness' would be stronger if the same backbone were evaluated both frozen and fine-tuned, and if the fine-tuned JOINT variant were compared with a simple ResNet-50 baseline under the same protocol.
minor comments (5)
  1. [Introduction] The phrase 'the then RAISE dataset' in the first paragraph of Section I appears to be a typo; it should likely read 'the RAISE dataset.'
  2. [Section V-A, Table II] The column header 'Performance (excluding col. 1 for rows 1–5)' is confusing. It would be clearer to state explicitly in the caption that rows 1–5 report performance after omitting the indicated feature category.
  3. [Section V-D] The weights for the overall JOINT model are reported as 0.769 for the rationality branch and 0.145 for the technical quality branch, but these do not sum to 1. Please clarify whether these are the exact weights from [10] or whether a normalization step is implied.
  4. [Section IV] The statement that 'even real images can be perceived as synthetic by human raters' is inferred from the overlapping MOS distributions in Fig. 5. This inference would be more direct if the paper reported the fraction of ratings below a threshold separately for real and AIGI subsets.
  5. [References] Reference [5] is typeset as 'A V A' with unusual spacing; please verify the rendering and use the standard 'AVA' formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper trains and evaluates supervised realness predictors on a held-out split of human MOS labels.

full rationale

The paper's derivation chain is empirical rather than definitional. The RAISE dataset is constructed from human subjective realness ratings (MOS), and each baseline model is trained on a 510-image partition and evaluated on a disjoint 90-image test partition (Section V). The reported correlations are computed against held-out human labels, so no 'prediction' is equivalent by construction to its training input. The JOINT model is an external baseline from Chen et al. with official weights, not a claim resting on the present authors' own prior work, and its adaptation involves training a new regressor on RAISE MOS labels. The skeptic's concern that mixed real/AIGI content may allow a model to succeed by source discrimination is a validity or interpretation risk, not a circularity, because the ground truth is perceived realness as rated by humans and the evaluation protocol does not substitute the source label for the human score. The dataset is self-referential only in the ordinary sense that 'realness' is operationalized through the collected ratings, which is the standard supervised-learning setup and does not make the evaluation circular. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors, and no known result is repackaged. Therefore the appropriate finding is no significant circularity.

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

No numbers were fitted ad hoc to support a derivation; the trained model weights are learned on the RAISE training split. The JOINT ensemble weights (0.769 and 0.145) are inherited from [10], not fitted here. The main assumptions are domain assumptions about the nature of realness and the reliability of the subjective ground truth.

assumptions (3)
  • domain assumption Perceptual realness is a distinct concept from technical quality and can be measured independently.
    The paper's framing separates realness from quality; Section III argues this using reference [10].
  • domain assumption MOS computed from at least 23 ratings after kurtosis-based outlier removal is a reliable ground truth.
    Section IV; reliability is asserted but no confidence intervals or inter-rater variability measures beyond SROCC are given.
  • domain assumption AIGIs sourced from Fake2M (Stable Diffusion v1.5) are representative of AI-generated images.
    Section III-A; if the generator distribution is narrow, baselines may not generalize to other generative models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RAISE: Realness Assessment for Image Synthesis and Evaluation." pith.science (2026). https://pith.science/paper/4LTHMA6Q

@misc{pith2026250519233,
  author       = {Pith},
  title        = {Pith review of: RAISE: Realness Assessment for Image Synthesis and Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4LTHMA6Q}},
  note         = {Machine review of arXiv:2505.19233}
}
read the original abstract

The rapid advancement of generative AI has enabled the creation of highly photorealistic visual content, offering practical substitutes for real images and videos in scenarios where acquiring real data is difficult or expensive. However, reliably substituting real visual content with AI-generated counterparts requires robust assessment of the perceived realness of AI-generated visual content, a challenging task due to its inherent subjective nature. To address this, we conducted a comprehensive human study evaluating the perceptual realness of both real and AI-generated images, resulting in a new dataset, containing images paired with subjective realness scores, introduced as RAISE in this paper. Further, we develop and train multiple models on RAISE to establish baselines for realness prediction. Our experimental results demonstrate that features derived from deep foundation vision models can effectively capture the subjective realness. RAISE thus provides a valuable resource for developing robust, objective models of perceptual realness assessment.

Figures

Figures reproduced from arXiv: 2505.19233 by the authors.

Figure 2
Figure 2. Rating screen shown to the participants of the human [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. Distribution of MOS ratings on the real and AIGI [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Images from the RAISE dataset corresponding to the [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Bar graphs showing the proportion of real and fake [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 7
Figure 7. Figure 7: Scatter plot of the predictions obtained by fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 25 canonical work pages

  1. [1]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

  2. [2]

    Hierarchical text-conditional image generation with clip latents,

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image generation with clip latents,” arXiv preprint arXiv:2204.06125, 2022

  3. [3]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,” IEEE Signal Processing Letters, vol. 20, no. 3, pp. 209–212, 2013

  4. [4]

    NIMA: Neural image assessment,

    H. Talebi and P. Milanfar, “NIMA: Neural image assessment,” IEEE transactions on image processing , vol. 27, no. 8, pp. 3998–4011, 2018

  5. [5]

    A V A: A large-scale database for aesthetic visual analysis,

    N. Murray, L. Marchesotti, and F. Perronnin, “A V A: A large-scale database for aesthetic visual analysis,” in IEEE conference on computer vision and pattern recognition , 2012, pp. 2408–2415

  6. [6]

    KonIQ-10k: An ecologi- cally valid database for deep learning of blind image quality assessment,

    V . Hosu, H. Lin, T. Sziranyi, and D. Saupe, “KonIQ-10k: An ecologi- cally valid database for deep learning of blind image quality assessment,” IEEE Transactions on Image Processing, vol. 29, pp. 4041–4056, 2020

  7. [7]

    A perceptual quality assessment exploration for AIGC images,

    Z. Zhang, C. Li, W. Sun, X. Liu, X. Min, and G. Zhai, “A perceptual quality assessment exploration for AIGC images,” in 2023 IEEE In- ternational Conference on Multimedia and Expo Workshops (ICMEW) , 2023, pp. 440–445

  8. [8]

    AGIQA-3k: An open database for AI-generated image quality assessment,

    C. Li, Z. Zhang, H. Wu, W. Sun, X. Min, X. Liu, G. Zhai, and W. Lin, “AGIQA-3k: An open database for AI-generated image quality assessment,” IEEE Transactions on Circuits and Systems for Video Technology, 2023

Show all 36 references
  1. [9]

    AIG- CIQA2023: A large-scale image quality assessment database for AI generated images: from the perspectives of quality, authenticity and correspondence,

    J. Wang, H. Duan, J. Liu, S. Chen, X. Min, and G. Zhai, “AIG- CIQA2023: A large-scale image quality assessment database for AI generated images: from the perspectives of quality, authenticity and correspondence,” in CAAI International Conference on Artificial Intel- ligence, 2...

  2. [10]

    Exploring the naturalness of AI-generated images,

    Z. Chen, W. Sun, H. Wu, Z. Zhang, J. Jia, Z. Ji, F. Sun, S. Jui, X. Min, G. Zhai et al., “Exploring the naturalness of AI-generated images,” arXiv preprint arXiv:2312.05476, 2023

  3. [11]

    AIGIQA-20k: A large database for AI- generated image quality assessment,

    C. Li, T. Kou, Y . Gao, Y . Cao, W. Sun, Z. Zhang, Y . Zhou, Z. Zhang, W. Zhang, H. Wu et al. , “AIGIQA-20k: A large database for AI- generated image quality assessment,” arXiv preprint arXiv:2404.03407 , vol. 2, no. 3, p. 5, 2024

  4. [12]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning , 2021, pp. 8748–8763

  5. [13]

    Improved techniques for training GANs,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training GANs,” Advances in neural information processing systems , vol. 29, 2016

  6. [14]

    GANs trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “GANs trained by a two time-scale update rule converge to a local nash equilibrium,” Advances in neural information processing systems , vol. 30, 2017

  7. [15]

    Demysti- fying MMD GANs,

    M. Bi ´nkowski, D. J. Sutherland, M. Arbel, and A. Gretton, “Demysti- fying MMD GANs,” arXiv preprint arXiv:1801.01401 , 2018

  8. [16]

    NTIRE 2024 quality assessment of ai-generated content challenge,

    X. Liu, X. Min, G. Zhai, C. Li, T. Kou, W. Sun, H. Wu, Y . Gao, Y . Cao, Z. Zhang et al., “NTIRE 2024 quality assessment of ai-generated content challenge,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 6337–6362

  9. [17]

    Subjective-aligned dataset and metric for text-to-video quality assessment,

    T. Kou, X. Liu, Z. Zhang, C. Li, H. Wu, X. Min, G. Zhai, and N. Liu, “Subjective-aligned dataset and metric for text-to-video quality assessment,” in Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 7793–7802

  10. [18]

    AiGC image quality assessment via image-prompt correspondence,

    F. Peng, H. Fu, A. Ming, C. Wang, H. Ma, S. He, Z. Dou, and S. Chen, “AiGC image quality assessment via image-prompt correspondence,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, vol. 6, 2024

  11. [19]

    AIGCc-VQA: A holistic perception metric for aigc video quality assessment,

    Y . Lu, X. Li, B. Li, Z. Yu, F. Guan, X. Wang, R. Liao, Y . Ye, and Z. Chen, “AIGCc-VQA: A holistic perception metric for aigc video quality assessment,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 6384–6394

  12. [20]

    Global-local image perceptual score (GLIPS): Evaluating photorealistic quality of ai-generated images,

    M. Aziz, U. Rehman, M. U. Danish, and K. Grolinger, “Global-local image perceptual score (GLIPS): Evaluating photorealistic quality of ai-generated images,” IEEE Transactions on Human-Machine Systems , 2025

  13. [21]

    Seeing is not always believing: Benchmarking human and model perception of ai-generated images,

    Z. Lu, D. Huang, L. Bai, J. Qu, C. Wu, X. Liu, and W. Ouyang, “Seeing is not always believing: Benchmarking human and model perception of ai-generated images,” Advances in Neural Information Processing Systems, vol. 36, pp. 25 435–25 447, 2023

  14. [22]

    KADID-10k: A large-scale artificially distorted iqa database,

    H. Lin, V . Hosu, and D. Saupe, “KADID-10k: A large-scale artificially distorted iqa database,” in 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX) , 2019, pp. 1–3

  15. [23]

    Most apparent distortion: full- reference image quality assessment and the role of strategy,

    E. C. Larson and D. M. Chandler, “Most apparent distortion: full- reference image quality assessment and the role of strategy,” Journal of Electronic Imaging , vol. 19, no. 1, p. 011006, 2010. [Online]. Available: https://doi.org/10.1117/1.3267105

  16. [24]

    Subjective video quality assessment methods for multimedia applications,

    ITU-T Recommendation P.910, “Subjective video quality assessment methods for multimedia applications,” 2008, https://www.itu.int/rec/ T-REC-P.910

  17. [25]

    Recommendation ITU-R BT.500-15: Methodologies for the subjective assessment of the quality of television images,

    International Telecommunication Union, Radiocommunication Sector (ITU-R), “Recommendation ITU-R BT.500-15: Methodologies for the subjective assessment of the quality of television images,” https://www. itu.int/rec/R-REC-BT.500-15-202305-I/en, 2023, accessed: 2025-05-04

  18. [26]

    Exploring varying color spaces through representative forgery learning to improve deepfake detection,

    M. A. Amin, Y . Hu, Y . Guan, and M. Z. Amin, “Exploring varying color spaces through representative forgery learning to improve deepfake detection,” Digital Signal Processing , vol. 147, p. 104426, 2024

  19. [27]

    Leveraging edges and optical flow on faces for deepfake detection,

    A. Chintha, A. Rao, S. Sohrawardi, K. Bhatt, M. Wright, and R. Ptucha, “Leveraging edges and optical flow on faces for deepfake detection,” in 2020 IEEE international joint conference on biometrics (IJCB) , 2020, pp. 1–10

  20. [28]

    DeepFake videos detection based on texture features

    B. Xu, J. Liu, J. Liang, W. Lu, and Y . Zhang, “DeepFake videos detection based on texture features.” Computers, Materials & Continua , vol. 68, no. 1, 2021

  21. [29]

    Textural features for image classification,

    R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural features for image classification,” IEEE Trans. Syst. Man Cybern., vol. SMC-3, no. 6, pp. 610–621, 1973

  22. [30]

    Histograms of oriented gradients for human detection,

    N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. (CVPR), vol. 1, 2005, pp. 886–893

  23. [31]

    Image feature detectors for deepfake video detection,

    F. F. Kharbat, T. Elamsy, A. Mahmoud, and R. Abdullah, “Image feature detectors for deepfake video detection,” in 2019 IEEE/ACS 16th International Conference on Computer Systems and Applications (AICCSA), 2019, pp. 1–4

  24. [32]

    Leveraging frequency analysis for deep fake image recogni- tion,

    J. Frank, T. Eisenhofer, L. Sch ¨onherr, A. Fischer, D. Kolossa, and T. Holz, “Leveraging frequency analysis for deep fake image recogni- tion,” in International conference on machine learning, 2020, pp. 3247– 3258

  25. [33]

    Dropout: a simple way to prevent neural networks from over- fitting,

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut- dinov, “Dropout: a simple way to prevent neural networks from over- fitting,” The journal of machine learning research , vol. 15, no. 1, pp. 1929–1958, 2014

  26. [34]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in International conference on machine learning , 2015, pp. 448–456

  27. [35]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 770–778

  28. [36]

    ImageNet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255

Pith tools

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