Pith. sign in

REVIEW 5 major objections 5 minor 76 references

Image Quality Assessment for Machines: Paradigm, Large-scale Database, and Models

T0 review · 5 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 Human-oriented image quality metrics cannot predict how degradation affects machine vision; the paper builds a 2.5-million-image machine-centric benchmark and a region-aware model that ranks machine-quality scores substantially better.

desk verdict A big, well-built MIQA benchmark and a sensible region-aware model; the broad anti-HVS conclusion goes beyond what the experiments actually show. read the letter →

arxiv 2508.19850 v1 pith:QGRHIV3C submitted 2025-08-27 cs.CV

Xiaoqi Wang , Yun Zhang , Weisi Lin This is my paper · ORCID
classification cs.CV
keywords machinevisionimagequalityassessmentmachine-centricconsistencyandaccuracyscoresdegradationrobustnessregion-awaremodeltransformerMIQD-2.5M
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 argues that image quality metrics built for human perception cannot tell whether a degraded image will still work for a machine vision system. To make the case, it constructs a database of 2.5 million degraded images labeled by how 75 vision models respond, in terms of both prediction consistency and task accuracy, across classification, detection, and segmentation. It then trains RA-MIQA, a region-aware model that scores images by their predicted impact on machine performance. On the new database, RA-MIQA ranks machine-centric quality substantially better than seven human-based quality metrics and five retrained backbones, with the largest gains on image classification. The paper also shows that background-dominated distortions and fine-grained tasks remain hard, and that a model trained for one task transfers poorly to a different one.

What carries the argument

The mean machine opinion score (MMOS) label is the ground truth that makes machine quality measurable: it aggregates consistency and accuracy responses of 75 pretrained vision models, weighted by their benchmark performance, for each degraded image. Spatial degradation modes—uniform (UD), ROI-dominated (ROI-DD), and background-dominated (BG-DD)—are generated using co-segmentation masks that split each image into task-relevant region and background, isolating where distortion hurts. The RA-MIQA architecture carries the argument: a frozen region encoder extracts a region token that joins a class token and patch tokens in a Transformer encoder, and a regression head fuses region and class repre

What would settle it

Take a held-out vision model that was not among the 75 used to build MMOS, degrade a set of images, and correlate RA-MIQA's predictions with that model's actual accuracy and consistency. If the correlation falls to near the level of human-based metrics, the ensemble-averaged labels were doing the predictive work rather than a general machine-quality signal. The paper's own cross-task numbers provide a concrete threshold: classification-trained RA-MIQA reaches only about 0.54 SRCC on detection and segmentation, so similarly weak transfer to an unseen task would falsify any claim that the model

Watch

Extended reading notes

Core claim

The central claim is that machine-centered image quality should be measured by how an image's degradations change what a vision system does, not by how they look to humans. The paper formalizes this as mean machine opinion score (MMOS), a weighted average over an ensemble of 75 vision models of a consistency score (agreement between predictions on the original and degraded image) and an accuracy score (agreement with ground truth), combined into a composite score. The database MIQD-2.5M supplies these labels for 2.5 million images spanning ten distortion types, five severity levels, and three spatial modes: uniform, ROI-dominated, and background-dominated distortions. The proposed RA-MIQA us

Load-bearing premise

The load-bearing premise is that the weighted average of responses from the 75 selected vision models defines machine image quality; if a target model outside that ensemble responds to degradations differently, the database labels and RA-MIQA will not predict its actual degradation impact.

Editorial extensions

If this is right

  • Human visual system metrics such as SSIM, LPIPS, and DISTS are not adequate proxies for machine perception; monitoring quality for deployed vision systems needs machine-centric scores.
  • Because MMOS labels are stable across random subsets of the 75 models (SRCC above 0.86 for classification and above 0.94 for detection and segmentation), ensemble-derived quality scores are a reproducible target for training.
  • RA-MIQA can rank degraded images by their expected impact on machine accuracy and consistency, enabling selection or restoration of images before they reach a downstream vision model.
  • MIQA is task-specific: classification-trained models transfer poorly to detection and segmentation (SRCC around 0.54), while detection and segmentation models transfer well to each other, so quality assessment should be tuned to task granularity.
  • Background-dominated and subtle distortions remain the hardest cases, pointing to the largest remaining gap in machine-centric quality assessment.

Reading between the lines

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

  • A testable extension is to use RA-MIQA's predicted scores to choose among preprocessing or restoration strategies per image, optimizing for downstream machine accuracy rather than pixel fidelity; the paper motivates this closed loop but does not demonstrate it.
  • The ensemble-based MMOS definition is tied to the 75-model training distribution, so for a new architecture or a fine-tuned edge model the safest deployment step would be to regenerate labels from that model or a small representative set, since cross-task and cross-model generalization is shown to be limited.
  • The consistency/accuracy split suggests operational monitoring should track both: the paper's own distributions show occasional cases of high consistency with low accuracy, meaning consistency alone could hide silent failures.
  • The region token's learned attention could be repurposed to localize which image regions a machine actually relies on, though the paper does not test this explainability use.
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

5 major / 5 minor

Summary. The paper proposes a machine-centric image quality assessment (MIQA) framework with consistency and accuracy labels, constructs MIQD-2.5M, a database of 2.5M degraded image-task samples over 75 models, 10 distortions, 3 spatial modes, and 3 tasks, and introduces RA-MIQA, a ViT-based regressor with a frozen region encoder. Extensive tables compare RA-MIQA with HVS metrics and retrained backbones, reporting SRCC gains of ~13% on classification and smaller gains on detection/segmentation. The authors conclude that HVS-based IQA metrics are inadequate for machine perception and that task/region-specific MIQA modeling is necessary.

Significance. If the claims hold, this is a substantial contribution: the first large-scale public MIQA database with multi-model, multi-task labels; a systematic benchmark showing HVS metrics correlate weakly with machine performance; a region-aware model with meaningful gains on classification; and a useful analysis of degradation-region/intensity interactions. Strengths include the 100-trial cross-model label validation (Table 4), the breadth of models/distortions, and the public release of model/code. The main caveat is that the evidence supports prediction of the 75-model ensemble score, not of arbitrary machine vision systems, so the broad 'fundamentally fail' conclusion requires additional held-out-model validation.

major comments (5)
  1. [§3.1.3, §4.2, Tables 5-6] The ground-truth MMOS is an ensemble average over the 75 models in Table 3 (Eqs. 2 and 4). All headline comparisons evaluate predictors against this aggregate label. The paper never validates that RA-MIQA predicts the quality experienced by a specific held-out model or an architecture outside the ensemble. The conclusion 'HVS-based IQA metrics fundamentally fail to capture machine perception requirements' is thus broader than the evidence. Please add a leave-one-model-out or held-out-architecture experiment in which per-model consistency/accuracy are correlated with RA-MIQA predictions, and report the distribution of per-model SRCC. This would directly test the machine-wide claim.
  2. [§3.1.3 Eq. (5), §4.2] The label-generation recipe is underspecified. Eqs. (2) and (4) use weights alpha_i, but the paper does not state how benchmark accuracies in Table 3 are converted into normalized alpha_i values, and the composite weights lambda1, lambda2 in Eq. (5) are never given. Without these values, MIQD-2.5M labels cannot be exactly reproduced and the composite-score results depend on undisclosed choices. Please report the exact alpha_i, lambda1/lambda2, and a sensitivity analysis showing how much SRCC changes under reasonable alternative weights.
  3. [§4.1, Eq. (6)] ROI masks are generated by co-segmentation [1] rather than from the task's ground-truth annotations. For detection and instance segmentation, the same generic co-segmentation mask may not cover all objects or may include non-task regions, so the ROI-DD vs BG-DD distinction may not correspond to 'task-relevant' vs 'task-irrelevant' regions. This directly affects the region-aware analysis and the RA-MIQA design. Please validate the masks against COCO bounding boxes/masks (e.g., IoU overlap statistics) or use task-specific ROIs for the fine-grained tasks.
  4. [§6.2, Tables 5-8] All reported correlations are from a single training run. The RA-MIQA gains over ViT-small on detection (SRCC +1.66%) and segmentation (+2.76%) are small, and without multiple seeds or confidence intervals it is unclear whether they are significant. Please report mean±std over at least 3-5 seeds for the main tables, and for the small gains include a significance test (e.g., paired bootstrap or Wilcoxon) or state if they are not significant.
  5. [§4.1] Classification source images are selected as the 'top 5 highest-rated images from each category' using HyperIQA and MANIQA, which are HVS-based NR-IQA models. This is exactly the family of metrics the paper argues is inadequate for machine perception. Such selection may bias the database toward images that are human-preferred but not necessarily machine-relevant, and it weakens the claim that the database is machine-centric. Please justify this choice or replace it with a task-agnostic criterion (e.g., resolution, clean acquisition) and discuss the potential bias.
minor comments (5)
  1. [Table 1] The 'Total Degraded Image = 2,500,000' appears to double-count: the 1,250,000 degraded COCO images are shared between detection and segmentation. If 2.5M refers to image-task samples rather than unique degraded images, please say so explicitly in the table and abstract to avoid inflating the apparent database size.
  2. [§6.1, Eq. (14)] The five-parameter logistic is fitted on the validation set and then PLCC/RMSE are computed on the same set. SRCC is rank-invariant and unaffected, but PLCC may be slightly optimistically biased. A cross-validated fitting or a statement about monotonic invariance would clarify.
  3. [Fig. 10 caption] The caption contains an apparent encoding artifact (a long run of '/uni...' tokens) that must be removed; the text should be restored to the intended description.
  4. [References, Table 3] Reference [4] is a duplicate of [3]; the citation for Cascade Mask R-CNN in Table 3 should point to the correct source. Also, the 'CO-DETR (ResNet-50)' row in Table 3 lacks a citation number.
  5. [Abstract / Conclusion] The phrase 'fundamentally fail' is strong given that only seven HVS metrics and synthetic distortions are tested. Consider softening to 'are substantially less correlated' or add evidence from additional HVS metrics and real-world corruptions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MMOS labels are an explicit operationalization, and RA-MIQA is trained and evaluated on held-out images against those labels in standard supervised fashion.

full rationale

The paper's derivation chain is not circular. The ground-truth machine quality (MMOS) is defined in Section 3.1.3 (Eqs. 2-5) as a weighted aggregate of consistency/accuracy responses over a fixed 75-model ensemble; this is an explicit operationalization, not a hidden restatement of the predictor's output. RA-MIQA (Section 5) maps the degraded image to a quality score, is trained with MSE loss on a training split, and is evaluated on held-out images (Section 6.1), so the reported SRCC/PLCC are out-of-sample correlations with independently computed labels. The HVS-baseline comparison is an empirical benchmark, not a logical consequence of the label definition. The cross-model label validation (Table 4) tests label stability across random model subsets; it does not fold the predictor's output into the label. The only fitted post-processing is the five-parameter logistic calibration (Eq. 14); because it is monotonic, SRCC (the headline metric) is invariant to it, and it is applied equally to all methods, so it does not create a fitted-input-called-prediction circularity. No load-bearing self-citation chain or imported uniqueness theorem is present. Concerns about whether the 75-model ensemble represents 'machine perception' generally, or about task-transfer limits (Figure 8), are scope and correctness issues, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The MMOS label definition, the ROI co-segmentation step, and the frozen region encoder are the main assumptions the central claim rests on. Free parameters include the ensemble weighting, the unspecified composite weights, the distortion severity mapping, and the validation-set-fitted logistic mapping.

free parameters (4)
  • MMOS model weights alpha_i = model benchmark metrics (top-1 acc, box AP, mask AP) from Table 3
    Weights in Eqs. 2 and 4 are taken from each model's reported performance; no sensitivity analysis of this choice is provided.
  • composite score weights lambda1, lambda2 = unspecified, presumably 0.5/0.5
    Default convex combination in Eq. 5; the paper never states the default values used in Tables 5-8.
  • distortion severity levels = 5 discrete levels per distortion type
    The mapping from severity level to physical parameters (noise sigma, blur radius, JPEG quality) is not specified in the paper; a re-implementation must guess these.
  • five-parameter logistic coefficients alpha1...alpha5 = fitted on the validation set
    Eq. 14 is fit to validation data before computing PLCC/RMSE; SRCC/KRCC are unaffected, but PLCC/RMSE values depend on this fitted mapping.
assumptions (5)
  • domain assumption The 75-model ensemble average is a valid definition of machine vision quality for the given task.
    Section 4.2 and 4.3; if the ensemble does not represent the target machine system, the MMOS ground truth is not meaningful.
  • domain assumption Co-segmentation based on Deep ViT features [1] correctly separates task-relevant ROIs from background.
    Section 4.1 step (b); errors in these masks mislabel all ROI-DD and BG-DD samples.
  • domain assumption Frozen region encoder [17] produces features informative for predicting MMOS.
    Section 5; RA-MIQA's region token is a single average-pooled vector from this encoder, so the performance gain depends on this encoder's quality.
  • domain assumption ImageNet pretrained weights are a valid initialization for MIQA regression.
    Section 6.2; all models start from ImageNet weights, which presumes transfer.
  • standard math Monotonic logistic regression does not change the rank comparison.
    Eq. 14 used for metric alignment; SRCC and KRCC are invariant under monotone transforms, so this is safe for those metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Image Quality Assessment for Machines: Paradigm, Large-scale Database, and Models." pith.science (2026). https://pith.science/paper/QGRHIV3C

@misc{pith2026250819850,
  author       = {Pith},
  title        = {Pith review of: Image Quality Assessment for Machines: Paradigm, Large-scale Database, and Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGRHIV3C}},
  note         = {Machine review of arXiv:2508.19850}
}
read the original abstract

Machine vision systems (MVS) are intrinsically vulnerable to performance degradation under adverse visual conditions. To address this, we propose a machine-centric image quality assessment (MIQA) framework that quantifies the impact of image degradations on MVS performance. We establish an MIQA paradigm encompassing the end-to-end assessment workflow. To support this, we construct a machine-centric image quality database (MIQD-2.5M), comprising 2.5 million samples that capture distinctive degradation responses in both consistency and accuracy metrics, spanning 75 vision models, 250 degradation types, and three representative vision tasks. We further propose a region-aware MIQA (RA-MIQA) model to evaluate MVS visual quality through fine-grained spatial degradation analysis. Extensive experiments benchmark the proposed RA-MIQA against seven human visual system (HVS)-based IQA metrics and five retrained classical backbones. Results demonstrate RA-MIQA's superior performance in multiple dimensions, e.g., achieving SRCC gains of 13.56% on consistency and 13.37% on accuracy for image classification, while also revealing task-specific degradation sensitivities. Critically, HVS-based metrics prove inadequate for MVS quality prediction, while even specialized MIQA models struggle with background degradations, accuracy-oriented estimation, and subtle distortions. This study can advance MVS reliability and establish foundations for machine-centric image processing and optimization. The model and code are available at: https://github.com/XiaoqiWang/MIQA.

Figures

Figures reproduced from arXiv: 2508.19850 by the authors.

Figure 1
Figure 1. The applications of MIQA in MVS: quality monitor [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The paradigm of MIQA. image block level. Dremin [14] proposed a CNN-based full￾reference MIQA model that quantifies the impact of vari￾ous image and video encoders on detection and recognition performance. Concurrent research by Li et al. [33], evalu￾ating over 30K distorted images across seven downstream tasks on 30 models, reveals a substantial discrepancy be￾tween human- and machine-oriented IQA. However, these m… view at source ↗
Figure 3
Figure 3. Pipeline of the MIQD-2.5M construction. 3.2.2 Task-Specific Response Spaces and Model Weighting The structure of response space Y is task-dependent: For classification, Y = {1, 2, . . . , C} for C discrete classes. In object detection, Y = {(b1, c1), . . . ,(bk, ck)} where bi denotes bounding box coordinates and ci the correspond￾ing class label. For segmentation, Y = {1, 2, . . . , C} H×W where H × W denotes spatia… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Density distribution of consistency and accuracy scores, where darker regions indicate higher sample concentration. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Quality distribution of images under different distor [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The distribution of MMOS values across different distortion regions and intensity levels in the three MIQA tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Structure of the proposed RA-MIQA model. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 9
Figure 9. Figure 9: The PLCC values of MIQA models across degrada [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Ablation study of the region encoder. Performance [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Sample illustration showcasing the corresponding [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 73 canonical work pages

  1. [33]

    Image Quality Assessment: From Human to Machine Preference

    Li, C., Tian, Y ., Ling, X., Zhang, Z., Duan, H., Wu, H., Jia, Z., Liu, X., Min, X., Lu, G., et al.: Image quality assessment: From human to machine preference. arXiv preprint arXiv:2503.10078 (2025)

    work page Pith review arXiv 2025

  2. [1]

    Proceedings of the European Confer- 14 Xiaoqi Wang et al

    Amir, S., Gandelsman, Y ., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. Proceedings of the European Confer- 14 Xiaoqi Wang et al. (i) Fog (BG-DD) (j) Snow (BG-DD) (e) Pixelate (ROI-DD) (f) JPEG (ROI-DD) (g) Motion Blur (ROI-DD) (h) Glass Blur (BG-DD) 28.40/0.977/0.115 0.634/0.405/0.492 0.617/0.301/0.459 24.88/0.919/0.302 0.303/0.20...

    work page 2022

  3. [2]

    In: Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2022)

    Beniwal, P., Mantini, P., Shah, S.K.: Image quality assessment us- ing deep features for object detection. In: Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2022)

    work page 2022

  4. [4]

    IEEE transactions on pattern analysis and machine intelligence 43(5), 1483–1498 (2019)

    Cai, Z., Vasconcelos, N.: Cascade r-cnn: High quality object de- tection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence 43(5), 1483–1498 (2019)

    work page 2019

  5. [5]

    In: European conference on computer vision, pp

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision, pp. 213–229. Springer (2020)

    work page 2020

  6. [6]

    In: Proceedings of the European conference on computer vision (ECCV), pp

    Chen, L.C., Zhu, Y ., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV), pp. 801–818 (2018)

    work page 2018

  7. [7]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Chen, Z., Xie, L., Niu, J., Liu, X., Wei, L., Tian, Q.: Visformer: The vision-friendly transformer. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 569–578 (2021)

    work page 2021

  8. [8]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmen- tation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 1290–1299 (2022)

    work page 2022

Show all 76 references
  1. [9]

    Advances in neural information processing systems 34, 3965–3977 (2021)

    Dai, Z., Liu, H., Le, Q.V ., Tan, M.: Coatnet: Marrying convolution and attention for all data sizes. Advances in neural information processing systems 34, 3965–3977 (2021)

    2021

  2. [10]

    Interna- tional Journal of Computer Vision 132(10), 4473–4509 (2024)

    Delussu, R., Putzu, L., Fumera, G.: Synthetic data for video surveillance applications of computer vision: A review. Interna- tional Journal of Computer Vision 132(10), 4473–4509 (2024)

    2024

  3. [11]

    In: 2009 IEEE conference on computer vision and pattern recognition, pp

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Ima- genet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition, pp. 248–

    2009

  4. [12]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2567–2581 (2022)

    Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assess- ment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2567–2581 (2022)

    2022

  5. [13]

    In: Interna- tional Conference on Learning Representations (2021)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Interna- tional Conference on...

    2021

  6. [14]

    In: International Conference on Pattern Recognition, pp

    Dremin, M., Kozhemyakov, K., Molodetskikh, I., Kirill, M., Ar- tur, S., Vatolin, D.: Machine vision-aware quality metrics for com- pressed image and video assessment. In: International Conference on Pattern Recognition, pp. 266–282. Springer (2024)

    2024

  7. [15]

    In: Proceedings of the IEEE/CVF international conference on computer vision, pp

    Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Center- net: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision, pp. 6569–6578 (2019) Image Quality Assessment for Machines: Paradigm, Large-scale Database...

    2019

  8. [16]

    In: 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pp

    Fezza, S.A., Bakhti, Y ., Hamidouche, W., D ´eforges, O.: Percep- tual evaluation of adversarial attacks for cnn-based image classifi- cation. In: 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pp. 1–6 (2019)

    2019

  9. [17]

    IEEE Transac- tions on Pattern Analysis and Machine Intelligence 45(6), 7457– 7476 (2023)

    Gao, S., Li, Z.Y ., Yang, M.H., Cheng, M.M., Han, J., Torr, P.: Large-scale unsupervised semantic segmentation. IEEE Transac- tions on Pattern Analysis and Machine Intelligence 45(6), 7457– 7476 (2023)

    2023

  10. [18]

    arXiv preprint arXiv:2107.08430 (2021)

    Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021)

    2021 arXiv

  11. [19]

    Advances in neural information processing systems 31 (2018)

    Geirhos, R., Temme, C.R., Rauber, J., Sch ¨utt, H.H., Bethge, M., Wichmann, F.A.: Generalisation in humans and deep neural net- works. Advances in neural information processing systems 31 (2018)

    2018

  12. [20]

    Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harness- ing adversarial examples. In: Y . Bengio, Y . LeCun (eds.) 3rd In- ternational Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceed- ings (2015)

    2015

  13. [21]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Graham, B., El-Nouby, A., Touvron, H., Stock, P., Joulin, A., J´egou, H., Douze, M.: Levit: a vision transformer in convnet’s clothing for faster inference. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 12239–12249 (2021)

    2021

  14. [22]

    IEEE transactions on pattern analysis and machine intelli- gence 29(4), 531–543 (2007)

    Grother, P., Tabassi, E.: Performance of biometric quality mea- sures. IEEE transactions on pattern analysis and machine intelli- gence 29(4), 531–543 (2007)

    2007

  15. [23]

    In: Pro- ceedings of the IEEE international conference on computer vision, pp

    He, K., Gkioxari, G., Doll ´ar, P., Girshick, R.: Mask r-cnn. In: Pro- ceedings of the IEEE international conference on computer vision, pp. 2961–2969 (2017)

    2017

  16. [24]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for im- age recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778 (2016)

    2016

  17. [25]

    Proceedings of the International Conference on Learning Representations (2019)

    Hendrycks, D., Dietterich, T.: Benchmarking neural network ro- bustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)

    2019

  18. [26]

    IEEE Transactions on Image Processing29, 4041–4056 (2020)

    Hosu, V ., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecolog- ically valid database for deep learning of blind image quality as- sessment. IEEE Transactions on Image Processing29, 4041–4056 (2020)

    2020

  19. [27]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pp

    Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pp. 7132–7141 (2018)

    2018

  20. [28]

    IEEE Transactions on Circuits and Systems for Video Technology 32(6), 3452–3461 (2022)

    Jin, J., Zhang, X., Fu, X., Zhang, H., Lin, W., Lou, J., Zhao, Y .: Just noticeable difference for deep machine vision. IEEE Transactions on Circuits and Systems for Video Technology 32(6), 3452–3461 (2022). DOI 10.1109/TCSVT.2021.3113572

    2022

  21. [29]

    URL https://github.com/ultralytics/ultralytics

    Jocher, G., Qiu, J., Chaurasia, A.: Ultralytics YOLO (2023). URL https://github.com/ultralytics/ultralytics

    2023

  22. [30]

    Inter- national journal of computer vision 129(2), 462–483 (2021)

    Kamann, C., Rother, C.: Benchmarking the robustness of semantic segmentation models with respect to common corruptions. Inter- national journal of computer vision 129(2), 462–483 (2021)

    2021

  23. [31]

    In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, pp

    Kirillov, A., Wu, Y ., He, K., Girshick, R.: Pointrend: Image seg- mentation as rendering. In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, pp. 9799–9808 (2020)

    2020

  24. [32]

    In: 2021 IEEE International Conference on Robotics and Automa- tion (ICRA), pp

    Lee, Y ., Jeon, J., Ko, Y ., Jeon, B., Jeon, M.: Task-driven deep im- age enhancement network for autonomous driving in bad weather. In: 2021 IEEE International Conference on Robotics and Automa- tion (ICRA), pp. 13746–13753 (2021)

    2021

  25. [34]

    IEEE Transactions on Pattern Analysis and Machine In- telligence 45(7), 8284–8295 (2023)

    Li, C., Zhou, H., Liu, Y ., Yang, C., Xie, Y ., Li, Z., Zhu, L.: Detection-friendly dehazing: Object detection in real-world hazy scenes. IEEE Transactions on Pattern Analysis and Machine In- telligence 45(7), 8284–8295 (2023)

    2023

  26. [35]

    In: European conference on computer vision, pp

    Li, Y ., Mao, H., Girshick, R., He, K.: Exploring plain vision trans- former backbones for object detection. In: European conference on computer vision, pp. 280–296. Springer (2022)

    2022

  27. [36]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Li, Y ., Wu, C.Y ., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transform- ers for classification and detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4794– 4804 (2022)

    2022

  28. [37]

    In: Computer vision–ECCV 2014: 13th European con- ference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pp

    Lin, T.Y ., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll´ar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer vision–ECCV 2014: 13th European con- ference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pp. 740–...

    2014

  29. [38]

    In: Proceedings of the 27th International Joint Conference on Ar- tificial Intelligence, pp

    Liu, D., Wen, B., Liu, X., Wang, Z., Huang, T.S.: When image denoising meets high-level vision tasks: a deep learning approach. In: Proceedings of the 27th International Joint Conference on Ar- tificial Intelligence, pp. 842–848 (2018)

    2018

  30. [39]

    Advances in neural information processing systems 34, 9204–9215 (2021)

    Liu, H., Dai, Z., So, D., Le, Q.V .: Pay attention to mlps. Advances in neural information processing systems 34, 9204–9215 (2021)

    2021

  31. [40]

    International Journal of Computer Vision 132(10), 4398–4416 (2024)

    Liu, J., Wang, Z., Ma, L., Fang, C., Bai, T., Zhang, X., Liu, J., Chen, Z.: Benchmarking object detection robustness against real-world corruptions. International Journal of Computer Vision 132(10), 4398–4416 (2024)

    2024

  32. [41]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Liu, Z., Lin, Y ., Cao, Y ., Hu, H., Wei, Y ., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9992–10002 (2021)

    2021

  33. [42]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Liu, Z., Mao, H., Wu, C.Y ., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11966–11976 (2022)

    2022

  34. [43]

    IEEE Transactions on Circuits and Systems for Video Technology 34(12), 13556– 13568 (2024)

    Lu, G., Ge, X., Zhong, T., Hu, Q., Geng, J.: Preprocessing en- hanced image compression for machine vision. IEEE Transactions on Circuits and Systems for Video Technology 34(12), 13556– 13568 (2024)

    2024

  35. [44]

    arXiv preprint arXiv:2212.07784 (2022)

    Lyu, C., Zhang, W., Huang, H., Zhou, Y ., Wang, Y ., Liu, Y ., Zhang, S., Chen, K.: Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784 (2022)

    2022 arXiv

  36. [45]

    In: European conference on computer vision, pp

    Maaz, M., Shaker, A., Cholakkal, H., Khan, S., Zamir, S.W., Anwer, R.M., Shahbaz Khan, F.: Edgenext: efficiently amalga- mated cnn-transformer architecture for mobile vision applications. In: European conference on computer vision, pp. 3–20. Springer (2022)

    2022

  37. [46]

    International Journal of Computer Vision 131(8), 1909–1963 (2023)

    Mao, J., Shi, S., Wang, X., Li, H.: 3d object detection for au- tonomous driving: A comprehensive survey. International Journal of Computer Vision 131(8), 1909–1963 (2023)

    1909

  38. [47]

    In: 2023 15th International Conference on Quality of Multimedia Experience (QoMEX), pp

    Marie, A., Desnos, K., Morin, L., Zhang, L.: Evaluation of im- age quality assessment metrics for semantic segmentation in a machine-to-machine communication scenario. In: 2023 15th International Conference on Quality of Multimedia Experience (QoMEX), pp. 1–6. IEEE (2023)

    2023

  39. [48]

    IEEE Transactions on Pattern Analysis and Machine Intel- ligence 43(4), 1239–1253 (2021)

    Pei, Y ., Huang, Y ., Zou, Q., Zhang, X., Wang, S.: Effects of image degradation and degradation removal to cnn-based image classifi- cation. IEEE Transactions on Pattern Analysis and Machine Intel- ligence 43(4), 1239–1253 (2021)

    2021

  40. [49]

    IEEE trans- actions on pattern analysis and machine intelligence 39(6), 1137– 1149 (2016)

    Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real- time object detection with region proposal networks. IEEE trans- actions on pattern analysis and machine intelligence 39(6), 1137– 1149 (2016)

    2016

  41. [50]

    In: European Conference on Computer Vision, pp

    Ren, Y ., Li, X., Li, B., Wang, X., Guo, M., Zhao, S., Zhang, L., Chen, Z.: Moe-diffir: Task-customized diffusion priors for univer- sal compressed image restoration. In: European Conference on Computer Vision, pp. 116–134. Springer (2025)

    2025

  42. [51]

    Radiology: Artificial Intelligence 2(1), e190015 (2020)

    Sabottke, C.F., Spieler, B.M.: The effect of image resolution on deep learning in radiography. Radiology: Artificial Intelligence 2(1), e190015 (2020)

    2020

  43. [52]

    ACM Computing Surveys (CSUR) 54(10s), 1–49 (2022) 16 Xiaoqi Wang et al

    Schlett, T., Rathgeb, C., Henniger, O., Galbally, J., Fierrez, J., Busch, C.: Face image quality assessment: A literature survey. ACM Computing Surveys (CSUR) 54(10s), 1–49 (2022) 16 Xiaoqi Wang et al

    2022

  44. [53]

    IEEE Transactions on Image Processing 19(6), 1427–1441 (2010)

    Seshadrinathan, K., Soundararajan, R., Bovik, A.C., Cormack, L.K.: Study of subjective and objective quality assessment of video. IEEE Transactions on Image Processing 19(6), 1427–1441 (2010)

    2010

  45. [54]

    arXiv preprint arXiv:1409.1556 (2014)

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  46. [55]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    Su, S., Yan, Q., Zhu, Y ., Zhang, C., Ge, X., Sun, J., Zhang, Y .: Blindly assess image quality in the wild guided by a self-adaptive hyper network. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    2020

  47. [56]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Sun, P., Zhang, R., Jiang, Y ., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., et al.: Sparse r-cnn: End-to-end object detection with learnable proposals. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14454–...

    2021

  48. [57]

    In: Proceedings of the AAAI conference on artificial intelli- gence, vol

    Szegedy, C., Ioffe, S., Vanhoucke, V ., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learn- ing. In: Proceedings of the AAAI conference on artificial intelli- gence, vol. 31 (2017)

    2017

  49. [58]

    IEEE Transactions on Image Processing 27(8), 3998–4011 (2018)

    Talebi, H., Milanfar, P.: Nima: Neural image assessment. IEEE Transactions on Image Processing 27(8), 3998–4011 (2018)

    2018

  50. [59]

    In: International conference on machine learning, pp

    Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convo- lutional neural networks. In: International conference on machine learning, pp. 6105–6114. PMLR (2019)

    2019

  51. [60]

    Advances in neural information processing systems 34, 24261–24272 (2021)

    Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An all-mlp architecture for vision. Advances in neural information processing systems 34, 24261–24272 (2021)

    2021

  52. [61]

    In: International conference on machine learning, pp

    Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., J´egou, H.: Training data-efficient image transformers & distilla- tion through attention. In: International conference on machine learning, pp. 10347–10357. PMLR (2021)

    2021

  53. [62]

    In: 2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pp

    Touvron, H., Cord, M., Sablayrolles, A., Synnaeve, G., J ´egou, H.: Going deeper with image transformers. In: 2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pp. 32–42 (2021)

    2021

  54. [63]

    Electronic Imaging 34, 1–6 (2022)

    Venkataramanan, A.K., Facktor, M., Gupta, P., Bovik, A.C.: As- sessing the impact of image quality on object-detection algo- rithms. Electronic Imaging 34, 1–6 (2022)

    2022

  55. [64]

    In: Proceedings of the IEEE/CVF international conference on computer vision, pp

    Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile back- bone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision, pp. 568–578 (2021)

    2021

  56. [65]

    IEEE Trans- actions on Image Processing 13(4), 600–612 (2004)

    Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality as- sessment: from error visibility to structural similarity. IEEE Trans- actions on Image Processing 13(4), 600–612 (2004)

    2004

  57. [66]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 16133–16142 (2023)

    2023

  58. [67]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    Xie, C., Zhang, Z., Zhou, Y ., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with in- put diversity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    2019

  59. [68]

    In: IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pp

    Xie, S., Girshick, R., Doll ´ar, P., Tu, Z., He, K.: Aggregated resid- ual transformations for deep neural networks. In: IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pp. 5987–5995 (2017)

    2017

  60. [69]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Xu, W., Xu, Y ., Chang, T., Tu, Z.: Co-scale conv-attentional image transformers. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9961–9970 (2021)

    2021

  61. [70]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pp

    Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y ., Cao, M., Wang, J., Yang, Y .: Maniqa: Multi-dimension attention network for no- reference image quality assessment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pp. 1191–1200 (2022)

    2022

  62. [71]

    Yu, W., Wang, X.: Mambaout: Do we really need mamba for vi- sion? In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    2025

  63. [72]

    In: The Eleventh International Conference on Learning Representations (2023)

    Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y .: Dino: Detr with improved denoising anchor boxes for end- to-end object detection. In: The Eleventh International Conference on Learning Representations (2023)

    2023

  64. [73]

    IEEE Transactions on Image Processing 23(10), 4270–4281 (2014)

    Zhang, L., Shen, Y ., Li, H.: Vsi: A visual saliency-induced index for perceptual image quality assessment. IEEE Transactions on Image Processing 23(10), 4270–4281 (2014)

    2014

  65. [74]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual met- ric. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    2018

  66. [75]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 7329–7338 (2023)

    2023

  67. [76]

    In: International Conference on Learning Representations (2021)

    Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021)

    2021

  68. [77]

    In: Proceedings of the IEEE/CVF international conference on computer vision, pp

    Zong, Z., Song, G., Liu, Y .: Detrs with collaborative hybrid assign- ments training. In: Proceedings of the IEEE/CVF international conference on computer vision, pp. 6748–6758 (2023)

    2023

Pith tools

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