Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Divisive Decisions: Improving Salience-Based Training for Generalization in Binary Classification Tasks

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Supervising both the true and false class activation maps during training improves generalization beyond true-class-only saliency guidance.

desk verdict A credible incremental idea for saliency-guided training, but the generalization claim needs significance testing and the abstract/conclusion are ahead of the data. read the letter →

arxiv 2507.17000 v1 pith:TF3XWPRM submitted 2025-07-22 cs.CV cs.LG

Jacob Piland , Chris Sweet , Adam Czajka This is my paper
classification cs.CVcs.LG
keywords saliency-guidedtrainingclassactivationmapsfalse-classCAMgeneralizationbinaryclassificationpresentationattackdetectionsyntheticfacepassivefooling
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

The paper argues that saliency-guided training for binary classifiers has been supervising only the true-class activation map, leaving the false-class map free to be misleading. It proposes three loss functions that bring both class maps into the objective, and reports that the contrastive variant achieves the best AUROC for iris presentation attack detection and improves synthetic face detection over the true-class-only baseline. It also introduces Difference Salience, the normalized difference of the two class maps, as a post-hoc visualization that recovers plausible decision-relevant features even in models trained to obfuscate their true-class CAMs.

What carries the argument

The central object is the class activation map (CAM) for a sample: the pixel-wise sum of the final-layer feature weights for a class times the activations, which feeds the logits. The load-bearing identity is that the difference between the true-class CAM and the false-class CAM is what decides a binary classification, so the difference map is a more faithful target for saliency supervision than the true-class CAM alone. The three proposed losses operationalize this by supervising the normalized CAM difference (Difference Salience), independently supervising both CAMs against human annotations and their inverse (Per-class Salience), or supervising the true CAM against human annotations and the false CAM against the inverse true CAM (Contrast Salience).

What would settle it

Train all three losses on a binary task using deliberately misleading human maps, such as image edges, and compare out-of-set AUROC with the true-CAM-only baseline: the authors' assumption predicts that when human attention is not aligned with discriminative cues, contrastive supervision should fail to improve or should degrade generalization.

Watch

Extended reading notes

Core claim

The paper's central claim is that in binary classification the quantity that determines the decision is the difference between the true- and false-class activation maps, so supervision that ignores the false-class map is incomplete. It proposes three training objectives built on this idea: Difference Salience matches the CAM difference to human saliency, Per-class Salience matches the true CAM to human saliency and the false CAM to its inverse, and Contrast Salience matches the true CAM to human saliency while driving the false CAM to the inverse of the true CAM. Across chest X-ray anomaly detection, synthetic face detection, and iris presentation attack detection, the paper finds that contrastive supervision improves out-of-set generalization beyond the baseline in the two generalization tasks, while the difference- and per-class variants produce a substantial gain on synthetic face detection.

Load-bearing premise

The methods assume that human saliency maps mark the features that genuinely separate the two classes, so that pushing the false-class CAM away from them produces decision-relevant divergence.

Editorial extensions

If this is right

  • If the central claim holds, saliency-guided training should routinely include false-class supervision rather than supervising only the true-class CAM.
  • Contrast Salience becomes the recommended variant for out-of-set generalization, with best iris PAD AUROC and improved synthetic face detection over baseline.
  • Difference Salience provides a post-hoc tool that resists passive fooling, so models whose true-class explanations are known to be untrustworthy can still be inspected.
  • The binary restriction matters: multi-class settings will need pairwise or embedding-based contrasts among multiple false classes.

Reading between the lines

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

  • An implicit testable corollary: Contrast Salience should also improve calibration or robustness to spurious correlations, since it actively discourages attention to non-discriminative regions.
  • If human saliency is noisy or misaligned, the contrastive losses may be less robust than the baseline because they add more dependence on the human map; one could measure sensitivity to annotation noise.
  • The success of Difference Salience as a visualization suggests that saliency evaluations should compare class-conditional maps, not just the true-class map.
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 extends saliency-guided training (the CYBORG baseline) from true-class CAMs only to losses that also constrain the false-class CAM in binary classification. Three losses are proposed—Difference Salience (Eq. 3), Per-class Salience (Eq. 4), and Contrast Salience (Eq. 5)—together with a post-hoc visualization, Difference Salience (Eq. 1), that subtracts the false-class CAM from the true-class CAM. The methods are evaluated on chest X-ray anomaly detection (in-set), iris presentation attack detection (out-of-set), and synthetic face detection (out-of-set), using AUROC over ten seeds. The authors report that the proposed losses improve generalization over the true-class-only baseline, with Contrast Salience highlighted as the most consistently beneficial method, and they qualitatively argue that Difference Salience reveals plausible features even in passively fooled models. The manuscript offers source code, model weights, and training configurations for reproducibility.

Significance. If the claimed generalization improvements were established, the paper would make a useful contribution to saliency-guided training by showing that false-class CAM supervision can add signal beyond true-class CAM alignment; the proposed Difference Salience visualization is also a sensible post-hoc diagnostic for passive-fooling settings. The experimental design is clear, the use of three domains with in-set and out-of-set protocols is appropriate, and the commitment to releasing code, weights, splits, and seeds is a genuine strength. However, the central empirical claim currently rests on AUROC differences that are small relative to the reported standard deviations, no significance testing is reported, all proposed losses underperform the baseline on the in-domain chest X-ray task, and the loss-weighting scheme introduces a confound between the baseline and the proposed methods. The idea is promising and the paper is reproducible, but the evidence as presented does not yet substantiate the abstract's unconditional 'improve generalization' statement.

major comments (4)
  1. [Abstract; Table 2; Sec. 8] The headline claim that the proposed methods 'improve generalization capabilities' over true-class-only saliency training is not supported by the aggregate results in Table 2. On the in-domain chest X-ray task, all three proposed losses are below the baseline (0.855–0.857 vs. 0.866±0.005). On Iris PAD, only Contrast Salience is above baseline, and only by 0.004 (0.790±0.089 vs. 0.786±0.091), while Difference and Per-class are below. On Synthetic Face, Contrast Salience improves by 0.007 (0.609±0.029 vs. 0.602±0.029), whereas the larger gains come from Difference and Per-class Salience (0.651 in both cases). The conclusion's statement that Contrast Salience 'achieves the best AUROC scores for iris PAD and improving on the baseline for synthetic face detection' is a selective reading of the same table, and the abstract's unconditional claim is an overstatement. The answer to the central question is at best task- and method-dependent, and the paper should be revised to state this explicitly.
  2. [Sec. 5.2; Table 2; Sec. 7] No statistical significance testing is reported anywhere in the paper. All comparisons are mean AUROC over ten seeds with standard deviations; for example, in Table 2 the Iris PAD Contrast Salience improvement is 0.004 with standard deviations of about 0.09, and the Synthetic Face Contrast Salience improvement is 0.007 with standard deviations of about 0.029. These differences are well within one standard deviation. Since all methods share the same DenseNet-121 backbone, the same data partitions, and the same classification component, paired per-seed comparisons are straightforward and would directly test the generalization claim. The limitations section (Sec. 7) discusses the human-saliency assumption but does not address the absence of significance testing. Without such tests, the reported improvements are not established.
  3. [Sec. 5.3; Eqs. (2)–(5)] The comparison between the baseline and the proposed methods is confounded by loss reweighting. The baseline in Eq. (2) uses α=β=0.5, while the proposed three-component losses in Eqs. (3)–(5) use α=β=γ=0.3. This means the classification cross-entropy term is weighted 0.5 in the baseline but only 0.3 in every proposed method. Any observed AUROC difference could therefore be due to the reduced classification weight rather than to the contrastive CAM mechanism. The manuscript does not report an ablation that holds the classification weight fixed (e.g., α=0.5 with β and γ chosen so that the total CAM supervision has comparable strength, or a sweep over weighting schemes). This must be addressed before the results can be attributed to the proposed losses.
  4. [Sec. 1.4; Sec. 5.1; Sec. 6.2–6.4] The research-question numbering in Sec. 5.1 is inconsistent with the definitions in Sec. 1.4. Item (b), 'supervising CAM difference,' is labeled as addressing RQ1, but RQ1 is defined as the qualitative passive-fooling visualization question; the quantitative difference-salience question is RQ2. Similarly, item (c) is labeled RQ2 but corresponds to RQ3 (per-class salience), and item (d) is labeled RQ3 but corresponds to RQ4 (contrast salience). The later sections (6.2–6.4) use the correct numbering, so this is a presentation error, but it makes the experimental design hard to follow and should be fixed.
minor comments (5)
  1. [Sec. 1.5] The phrase 'demonstrate it's value' contains a typo; it should be 'demonstrate its value.'
  2. [Eq. (1)] The text says 'remaps x to a unite interval'; the intended term is 'unit interval.'
  3. [References] References [29] and [30] are the same paper (Trokielewicz et al. on iris recognition reliability with ocular pathologies) and should be merged into a single entry.
  4. [Sec. 6.1] The answer to RQ1 is based entirely on visual inspection of six illustrative examples in Fig. 1. Since the question is qualitative this is acceptable, but the claim that Difference Salience 'reveals new and plausible features' could be strengthened by a quantitative saliency-quality metric or by reporting inter-subset consistency.
  5. [Sec. 5.4; Table 1] The manuscript does not describe how the human saliency maps were obtained for each dataset (image annotations vs. eye tracking, and their sources). Because the human maps are the supervision signal in all proposed losses, a brief description or pointer for each domain would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical AUROC comparisons on held-out partitions, and the proposed losses encode the hypothesis rather than entailing the measured outcome.

full rationale

The paper's central claims are empirical: that the three proposed losses (Eqs. 3-5) improve AUROC on held-out partitions relative to the CYBORG baseline (Eq. 2). The loss functions directly encode the hypothesis that true- and false-class CAMs should diverge on human-important features, but the evaluation metric (AUROC on unseen test partitions, Table 2) is not optimized by the losses and is not derived from them, so the reported gains are not equivalent to the training objective by construction. No parameter is fitted to the test set; the baseline comes from prior published CYBORG work, and the paper explicitly reuses fixed train/test partitions from that work for comparability, which is independent evidence rather than a self-citation chain. The RQ1 qualitative claim about Difference Salience is definitionally related to the logit difference (Eq. 1 and the text 'it is their difference that decides an input's classification label'), but the paper frames it as a proposed visualization and supports it with qualitative examples, not as a theorem; any lack of novelty would be a contribution concern, not circularity. The acknowledged limitation in Sec. 7 (human salience maps may not correspond to discriminative cues) is a load-bearing assumption about external annotations, not a circular step. The absence of significance testing for AUROC gaps that are within one standard deviation is a statistical robustness concern, but the hard-rule standard for circularity (exhibiting Eq. X = Eq. Y by construction, or a fitted parameter renamed as a prediction) is not met.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a few hand-chosen loss weights and on the assumption that human attention maps mark the features that truly separate the classes. No new physical or conceptual entities are introduced; the 'Difference Salience' map is a derived visualization, not an invented entity. The paper relies on established CAM machinery and on the prior CYBORG baseline.

free parameters (3)
  • alpha (classification loss weight) = 0.5 (two-part loss) or 0.3 (three-part loss)
    Hand-chosen equal weighting in all experiments; no sensitivity analysis is reported (Sec. 5.3).
  • beta (human saliency MSE weight) = 0.5 or 0.3
    Set equal to alpha without tuning; the central claim's magnitude likely depends on this ratio.
  • gamma (false-class supervision weight) = 0.3
    Only used in the three-part losses (Eqs. 4-5); hand-selected, no ablation over weights.
assumptions (4)
  • domain assumption Human saliency maps indicate classification-relevant features
    Invoked in Eqs. (2)-(5); the authors explicitly flag it as an assumption in Sec. 7.
  • domain assumption MSE between normalized saliency maps is a suitable supervision objective
    Used in all losses; no justification is given beyond prior practice (CYBORG).
  • domain assumption The class activation map is computed as the pixel-wise product of final-layer weights and activations, and comparing CAMs of different classes via subtraction is meaningful
    Defined in Sec. 3.1 as the input to the final linear layer; the Difference Salience and all contrast losses depend on this definition.
  • domain assumption A single ImageNet-pretrained DenseNet-121 can serve as the backbone for all three domains
    Sec. 5.3 instantiates all models this way; no domain-specific architecture adaptation is explored.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Divisive Decisions: Improving Salience-Based Training for Generalization in Binary Classification Tasks." pith.science (2026). https://pith.science/paper/TF3XWPRM

@misc{pith2026250717000,
  author       = {Pith},
  title        = {Pith review of: Divisive Decisions: Improving Salience-Based Training for Generalization in Binary Classification Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TF3XWPRM}},
  note         = {Machine review of arXiv:2507.17000}
}
read the original abstract

Existing saliency-guided training approaches improve model generalization by incorporating a loss term that compares the model's class activation map (CAM) for a sample's true-class ({\it i.e.}, correct-label class) against a human reference saliency map. However, prior work has ignored the false-class CAM(s), that is the model's saliency obtained for incorrect-label class. We hypothesize that in binary tasks the true and false CAMs should diverge on the important classification features identified by humans (and reflected in human saliency maps). We use this hypothesis to motivate three new saliency-guided training methods incorporating both true- and false-class model's CAM into the training strategy and a novel post-hoc tool for identifying important features. We evaluate all introduced methods on several diverse binary close-set and open-set classification tasks, including synthetic face detection, biometric presentation attack detection, and classification of anomalies in chest X-ray scans, and find that the proposed methods improve generalization capabilities of deep learning models over traditional (true-class CAM only) saliency-guided training approaches. We offer source codes and model weights\footnote{GitHub repository link removed to preserve anonymity} to support reproducible research.

Figures

Figures reproduced from arXiv: 2507.17000 by the authors.

Figure 1
Figure 1. Illustrative examples from three binary class datasets of misleading CAMs produced by models training with passive fooling and [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 33 canonical work pages

  1. [1]

    Ac- cessed: 03-12-2021

    Chinese academy of sciences institute of automation. Ac- cessed: 03-12-2021. 4, 5

    work page 2021

  2. [2]

    SREFI: Synthesis of realistic example face images

    Sandipan Banerjee, John S Bernhard, Walter J Scheirer, Kevin W Bowyer, and Patrick J Flynn. SREFI: Synthesis of realistic example face images. In IEEE Int. Joint Conf. on Biometrics (IJCB), pages 37–45. IEEE, 2017. 4, 5

    work page 2017

  3. [3]

    Human- aided saliency maps improve generalization of deep learning

    Aidan Boyd, Kevin W Bowyer, and Adam Czajka. Human- aided saliency maps improve generalization of deep learning. In Proceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision , pages 2735–2744, 2022. 2, 4

    work page 2022

  4. [4]

    Cyborg: Blending human saliency into the loss im- proves deep learning-based synthetic face detection

    Aidan Boyd, Patrick Tinsley, Kevin W Bowyer, and Adam Czajka. Cyborg: Blending human saliency into the loss im- proves deep learning-based synthetic face detection. In Pro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 6108–6117, 2023. 1, 2, 4

    work page 2023

  5. [5]

    Grad-cam++: General- ized gradient-based visual explanations for deep convolu- tional networks

    Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: General- ized gradient-based visual explanations for deep convolu- tional networks. In 2018 IEEE winter conference on appli- cations of computer vision (WACV) , pages 839–847. IEEE,

    2018

  6. [6]

    StarGAN: Unified Gen- erative Adversarial Networks for Multi-Domain Image-to- Image Translation

    Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified Gen- erative Adversarial Networks for Multi-Domain Image-to- Image Translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 5

    work page 2018

  7. [7]

    Use hirescam in- stead of grad-cam for faithful explanations of convolutional neural networks

    Rachel Lea Draelos and Lawrence Carin. Use hirescam in- stead of grad-cam for faithful explanations of convolutional neural networks. arXiv preprint arXiv:2011.08891, 2020. 2

    arXiv 2011

  8. [8]

    Axiom-based grad-cam: To- wards accurate visualization and explanation of cnns

    Ruigang Fu, Qingyong Hu, Xiaohu Dong, Yulan Guo, Yinghui Gao, and Biao Li. Axiom-based grad-cam: To- wards accurate visualization and explanation of cnns. arXiv preprint arXiv:2008.02312, 2020. 2

    arXiv 2008

Show all 37 references
  1. [9]

    Iris liveness detection based on quality re- lated features

    Javier Galbally, Jaime Ortiz-Lopez, Julian Fierrez, and Javier Ortega-Garcia. Iris liveness detection based on quality re- lated features. In 2012 5th IAPR International Conference on Biometrics (ICB), pages 271–276. IEEE, 2012. 4, 5

    2012

  2. [10]

    ProGAN: Net- work Embedding via Proximity Generative Adversarial Net- work

    Hongchang Gao, Jian Pei, and Heng Huang. ProGAN: Net- work Embedding via Proximity Generative Adversarial Net- work. In Proceedings of the 25th ACM SIGKDD Interna- tional Conference on Knowledge Discovery & Data Mining, page 1308–1316, New York, NY , USA, 2019. Association for...

    2019

  3. [11]

    Fooling neu- ral network interpretations via adversarial model manipula- tion

    Juyeon Heo, Sunghwan Joo, and Taesup Moon. Fooling neu- ral network interpretations via adversarial model manipula- tion. Advances in neural information processing systems, 32,

  4. [12]

    Weinberger

    Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kil- ian Q. Weinberger. Densely connected convolutional net- works. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4

    2017

  5. [13]

    Mimic-cxr, a de- identified publicly available database of chest radiographs with free-text reports

    Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de- identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 5

    2019

  6. [14]

    Progressive growing of gans for improved quality, stability, and variation, 2017

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2017. 4, 5

    2017

  7. [15]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 4

    2019

  8. [16]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4

    2019

  9. [17]

    Training generative adver- sarial networks with limited data

    Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adver- sarial networks with limited data. In Proc. NeurIPS, 2020. 4

    2020

  10. [18]

    Analyzing and improv- ing the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improv- ing the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119, 2020. 4

    2020

  11. [19]

    Alias-free generative adversarial networks

    Tero Karras, Miika Aittala, Samuli Laine, Erik H ¨ark¨onen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 4, 5

    2021

  12. [20]

    Revisiting iris recognition with color cosmetic con- tact lenses

    Naman Kohli, Daksha Yadav, Mayank Vatsa, and Richa Singh. Revisiting iris recognition with color cosmetic con- tact lenses. In 2013 International Conference on Biometrics (ICB), pages 1–7. IEEE, 2013. 4, 5

    2013

  13. [21]

    Detecting medley of iris spoofing attacks using desist

    Naman Kohli, Daksha Yadav, Mayank Vatsa, Richa Singh, and Afzel Noore. Detecting medley of iris spoofing attacks using desist. In 2016 IEEE 8th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1–6. IEEE, 2016. 4

    2016

  14. [22]

    Multifeature-based fake iris detection method

    Sung Joo Lee, Kang Ryoung Park, Youn Joo Lee, Kwanghyuk Bae, and Jaihie Kim. Multifeature-based fake iris detection method. Optical Engineering, 46(12):127204– 127204, 2007. 4, 5

    2007

  15. [23]

    Eigen-cam: Class activation map using principal compo- nents

    Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal compo- nents. In 2020 international joint conference on neural net- works (IJCNN), pages 1–7. IEEE, 2020. 2

    2020

  16. [24]

    Warsaw datasets web- page

    Warsaw University of Technology. Warsaw datasets web- page. http://zbum.ia.pw.edu.pl/EN/node/46 ,

  17. [25]

    Lessons from collecting a million biometric samples

    P Jonathon Phillips, Patrick J Flynn, and Kevin W Bowyer. Lessons from collecting a million biometric samples. Image and Vision Computing, 58:96–107, 2017. 4, 5

    2017

  18. [26]

    Ablation-cam: Visual explanations for deep convolutional network via gradient- free localization

    Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient- free localization. In proceedings of the IEEE/CVF winter conference on applications of computer vision , pages 983– 991, 2020. 2

    2020

  19. [27]

    Eye movement- driven defense against iris print-attacks

    Ioannis Rigas and Oleg V Komogortsev. Eye movement- driven defense against iris print-attacks. Pattern Recognition Letters, 68:316–326, 2015. 4, 5 8

    2015

  20. [28]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE in- ternational conference on computer vision , pages 618–626,

  21. [29]

    Assessment of iris recognition reliability for eyes af- fected by ocular pathologies

    Mateusz Trokielewicz, Adam Czajka, and Piotr Maciejew- icz. Assessment of iris recognition reliability for eyes af- fected by ocular pathologies. In 2015 IEEE 7th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1–6. IEEE, 2015. 4, 5

    2015

  22. [30]

    Assessment of iris recognition reliability for eyes af- fected by ocular pathologies

    Mateusz Trokielewicz, Adam Czajka, and Piotr Maciejew- icz. Assessment of iris recognition reliability for eyes af- fected by ocular pathologies. In 2015 IEEE 7th International Conference on Biometrics Theory, Applications and Systems (BTAS), pages 1–6. IEEE, 2015

    2015

  23. [31]

    Post-mortem iris recognition with deep-learning-based image segmentation

    Mateusz Trokielewicz, Adam Czajka, and Piotr Maciejew- icz. Post-mortem iris recognition with deep-learning-based image segmentation. Image and Vision Computing , 94: 103866, 2020. 4, 5

    2020

  24. [32]

    Score-cam: Score-weighted visual explanations for convolutional neural networks

    Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops ...

    2020

  25. [33]

    Synthesis of large realistic iris databases using patch-based sampling

    Zhuoshi Wei, Tieniu Tan, and Zhenan Sun. Synthesis of large realistic iris databases using patch-based sampling. In 2008 19th International Conference on Pattern Recognition, pages 1–4. IEEE, 2008. 4, 5

    2008

  26. [34]

    Livdet iris 2017-iris liveness detection competition 2017

    David Yambay, Benedict Becker, Naman Kohli, Daksha Ya- dav, Adam Czajka, Kevin W Bowyer, Stephanie Schuckers, Richa Singh, Mayank Vatsa, Afzel Noore, et al. Livdet iris 2017-iris liveness detection competition 2017. . 4

    2017

  27. [35]

    Livdet-iris 2015–iris liveness detection com- petition 2015

    David Yambay, Brian Walczak, Stephanie Schuckers, and Adam Czajka. Livdet-iris 2015–iris liveness detection com- petition 2015. . 4, 5

    2015

  28. [36]

    Learning deep features for discrimi- native localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discrimi- native localization. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 2921–2929,

    2016

  29. [37]

    Learning deep features for discrimina- tive localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discrimina- tive localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921–2929,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.