arxiv: 2604.22990 · v2 · submitted 2026-04-24 · 💻 cs.CV · cs.AI

Recognition: unknown

Hard to See, Hard to Label: Generative and Symbolic Acquisition for Subtle Visual Phenomena

Renjith Prasad , Rishabh Sharma , Andrew E. Shao , Annmary Justine Koomthanam , Shreyas Kulkarni , Suparna Bhattacharya , Martin Foltin , Amit Sheth

show 3 more authors

David Orozco Matthew Quinn Brian Sammuli

Authors on Pith no claims yet

Pith reviewed 2026-05-08 12:13 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords active learningobject detectionsubtle anomaliesdiffusion modelssemantic coveragedefect detectionlabel efficiencyrare class retrieval

0 comments

The pith

GSAL combines diffusion difficulty scoring with semantic concept graphs to acquire subtle visual anomalies more effectively in active learning.

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

The paper targets the problem of active learning missing subtle, low-contrast visual anomalies such as hairline cracks or voids because uncertainty and diversity methods favor common patterns. GSAL scores candidate images and proposals for visual difficulty using reconstruction discrepancy and denoising variability from diffusion models, then balances this against a three-level concept graph that enforces coverage of underrepresented semantic regions. Experiments on thin-film defect data plus Pascal VOC and MS COCO show higher label efficiency and better rare-class retrieval than uncertainty, diversity, or hybrid baselines. A sympathetic reader would care because the method supplies both a practical acquisition rule and interpretable rationales for why certain hard examples are chosen. If the balance works, it implies active learning can systematically reach sparse but important visual phenomena without extra human effort.

Core claim

GSAL is an active learning framework for object detection that combines a diffusion-based difficulty signal with a hierarchical semantic coverage prior. The diffusion component scores images and proposals using reconstruction discrepancy and denoising variability to prioritize visually atypical or ambiguous examples. The semantic component organizes candidates in a three-level concept graph to promote coverage of underrepresented regions while supplying interpretable acquisition rationales. By balancing the two, GSAL improves retrieval of subtle and rare targets that uncertainty-only selection often misses, with consistent gains in label efficiency on a proprietary thin-film defect dataset,

What carries the argument

GSAL's dual acquisition mechanism: diffusion-derived visual difficulty (reconstruction discrepancy plus denoising variability) paired with a three-level concept graph that enforces semantic coverage.

If this is right

GSAL yields higher label efficiency than uncertainty-, diversity-, and hybrid baselines on industrial defect data and on Pascal VOC and MS COCO.
The method retrieves more rare classes that are typically overlooked by uncertainty-only selection.
Acquisition decisions become interpretable through the semantic concept graph rationales.
Diffusion scoring prevents repeated selection of hard examples that still lie inside dominant semantic modes.
The framework applies directly to object detection tasks where anomalies are both low-prevalence and visually ambiguous.

Where Pith is reading between the lines

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

The same difficulty-plus-coverage logic could be tested on medical imaging or autonomous-driving edge-case datasets where subtle anomalies also matter.
If the concept graph can be learned automatically rather than hand-built, the method would require less domain expertise to deploy.
Replacing the diffusion model with other generative models would test whether the difficulty signal is specific to diffusion or more general.
In high-volume industrial inspection, the reported gains in rare-class retrieval could translate into measurable reductions in missed defects and therefore in warranty or safety costs.

Load-bearing premise

The diffusion signals must identify subtle anomalies without being overwhelmed by unrelated image properties, and the concept graph must be constructible in a way that genuinely covers rare semantic regions without adding its own selection biases.

What would settle it

On a dataset containing a known set of subtle anomalies with ground-truth labels, compare the fraction of those anomalies acquired by GSAL versus pure uncertainty sampling after a fixed labeling budget; a clear drop in recall for the subtle set would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.22990 by Amit Sheth, Andrew E. Shao, Annmary Justine Koomthanam, Brian Sammuli, David Orozco, Martin Foltin, Matthew Quinn, Renjith Prasad, Rishabh Sharma, Shreyas Kulkarni, Suparna Bhattacharya.

**Figure 1.** Figure 1: Qualitative examples from the industrial thin-film defect dataset. Representative defect types include inclusion, drill hole, Petal, and void. These defects are often small, weakly contrasted, and visually confounded with background texture, making them easy to overlook under standard active learning heuristics. 1. Introduction Current AL methods for object detection rely on discriminative uncertainty [… view at source ↗

**Figure 2.** Figure 2: Overview of GSAL. Given an unlabeled image, the detector first generates region proposals. We then compute diffusion-based difficulty scores at two levels: on the full image to capture global scene ambiguity, and on each proposal to capture localized subtle anomalies that may be diluted at image level. In parallel, each proposal receives a rarity-aware bonus from a three-level concept graph that tracks und… view at source ↗

**Figure 3.** Figure 3: Representative concept graph for the industrial thinfilm defect dataset. The figure illustrates how defect types such as void and drill hole are organized across three semantic levels in GSAL: fine labels, coarse morphological groups, and nonexclusive abstract attributes. This hierarchy makes semantic underrepresentation explicit at multiple granularities and supports interpretable acquisition rationa… view at source ↗

**Figure 4.** Figure 4: GSAL acquisition procedure for one active-learning round. 80/10/10 into train/val/test. The dataset contains four defect types, inclusions, voids, petals, and drill holes that are small in scale, low-contrast with surrounding material, and structurally similar to background noise, making annotation particularly costly and label efficiency critical. Ground-truth labels are binary (defect present vs. no de… view at source ↗

read the original abstract

Subtle visual anomalies such as hairline cracks, sub-millimeter voids, and low-contrast inclusions are structurally atypical yet visually ambiguous, making them both difficult to annotate and easy to overlook during active learning. Standard acquisition heuristics based on discriminative uncertainty or feature diversity often overselect dominant patterns while underexploring sparse yet important regions of the data space. This failure mode is especially severe in industrial defect inspection, where anomalies may be both low-prevalence and difficult to distinguish from surrounding structure. To resolve this, we propose GSAL, an active learning framework for object detection that combines a diffusion-based difficulty signal with a hierarchical semantic coverage prior. The diffusion component scores images and proposals using reconstruction discrepancy and denoising variability, prioritizing visually atypical or ambiguous examples. However, diffusion alone does not prevent acquisition from repeatedly favoring hard samples within dominant semantic modes. The semantic component therefore organizes candidate samples in a three-level concept graph and promotes coverage of underrepresented semantic regions while providing interpretable acquisition rationales. By balancing visual difficulty with semantic coverage, GSAL improves retrieval of subtle and rare targets that are often missed by uncertainty-only selection. Experiments on a proprietary thin-film defect, Pascal VOC and MS COCO dataset show consistent gains in label efficiency and rare-class retrieval over uncertainty-, diversity-, and hybrid-based baselines

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

GSAL pairs diffusion difficulty scoring with a three-level concept graph to target subtle rare anomalies in active learning, but the abstract gives almost no numbers or details to judge the gains.

read the letter

The main thing here is that the paper proposes GSAL, which scores candidate samples for visual difficulty using diffusion reconstruction error and denoising variability, then balances that against coverage in a three-level semantic concept graph so that rare or atypical categories do not get ignored. This is aimed squarely at object detection for low-prevalence defects like hairline cracks or low-contrast inclusions where standard uncertainty sampling fails by over-picking easy dominant patterns. The authors test on a proprietary thin-film dataset plus Pascal VOC and MS COCO and claim better label efficiency and rare-class retrieval than uncertainty, diversity, and hybrid baselines. The logic is internally consistent and the problem statement matches a real industrial pain point. Using public datasets alongside the proprietary one is a positive step for partial verification. The concept graph is also presented as giving interpretable acquisition reasons, which could be practically useful. The soft spots are straightforward. The abstract contains no quantitative results, error bars, ablation numbers, or implementation specifics on how the two signals are fused or how the graph is constructed from data. Without those, it is impossible to tell whether the claimed improvements are large enough to matter or whether the diffusion signal is actually picking subtlety rather than other image properties. The proprietary dataset also blocks easy reproduction. This work is for people building active learning pipelines for defect detection or other imbalanced visual tasks. A practitioner in industrial CV might pick up the high-level idea and try to adapt the balancing approach, but a broader researcher would probably wait for stronger numbers before citing or extending it. I would send it to peer review. The framing is coherent, the motivation is clear, and the combination is not a trivial restatement of prior active learning work, so referees can usefully ask for the missing details and check the experiments.

Referee Report

3 major / 1 minor

Summary. The paper introduces GSAL, an active learning framework for object detection that integrates a diffusion-based difficulty signal (reconstruction discrepancy and denoising variability) with a three-level hierarchical concept graph for semantic coverage. It claims this combination balances visual atypicality with underrepresented semantic regions to improve retrieval of subtle and rare targets (e.g., hairline cracks, low-contrast inclusions) that uncertainty- or diversity-only methods miss. Experiments on a proprietary thin-film defect dataset plus Pascal VOC and MS COCO report consistent gains in label efficiency and rare-class retrieval over uncertainty, diversity, and hybrid baselines.

Significance. If the results hold with full quantitative support and public verification, the work could meaningfully advance active learning for domains with low-prevalence, visually ambiguous anomalies such as industrial inspection. The dual generative-symbolic design and emphasis on interpretable acquisition rationales are strengths that distinguish it from purely uncertainty-driven approaches. The paper correctly identifies a failure mode of standard heuristics and proposes a coherent mitigation.

major comments (3)

[Abstract] Abstract: the central claim of 'consistent gains in label efficiency and rare-class retrieval' is asserted without any reported numbers, error bars, statistical tests, or ablation details on how the diffusion signal is combined with the concept-graph prior. This is load-bearing for the empirical contribution.
[Experiments] Experiments section: reliance on a proprietary thin-film defect dataset prevents independent replication of the industrial-defect results that motivate the method; this blocks verification of the key practical claim.
[Method] Method description: no explicit formulation or pseudocode is given for the joint acquisition function that trades off diffusion difficulty against semantic coverage, nor for construction of the three-level concept graph. This leaves open whether the reported improvement reduces to an ad-hoc weighting parameter.

minor comments (1)

[Figures/Tables] Figure captions and table headers should explicitly state the exact metrics (e.g., mAP@0.5, rare-class recall) and number of runs used for the reported gains.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below, indicating the revisions we will make to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of 'consistent gains in label efficiency and rare-class retrieval' is asserted without any reported numbers, error bars, statistical tests, or ablation details on how the diffusion signal is combined with the concept-graph prior. This is load-bearing for the empirical contribution.

Authors: We agree that the abstract would benefit from more concrete quantitative support. In the revised manuscript we will update the abstract to include specific metrics (e.g., relative gains in label efficiency and rare-class recall on the evaluated datasets), reference the corresponding tables, and briefly note the ablation results that isolate the contribution of the diffusion signal versus the semantic coverage prior. revision: yes
Referee: [Experiments] Experiments section: reliance on a proprietary thin-film defect dataset prevents independent replication of the industrial-defect results that motivate the method; this blocks verification of the key practical claim.

Authors: We acknowledge that the proprietary thin-film dataset limits full independent replication of the motivating industrial results. To mitigate this, the revision will place greater emphasis on the fully reproducible results obtained on Pascal VOC and MS COCO, include additional analysis of generalization across these public benchmarks, and release the complete implementation code so that the algorithmic components and public-dataset experiments can be reproduced exactly. revision: partial
Referee: [Method] Method description: no explicit formulation or pseudocode is given for the joint acquisition function that trades off diffusion difficulty against semantic coverage, nor for construction of the three-level concept graph. This leaves open whether the reported improvement reduces to an ad-hoc weighting parameter.

Authors: We thank the referee for highlighting this omission. The revised manuscript will contain the explicit mathematical formulation of the joint acquisition function (including the precise weighting between diffusion difficulty and semantic coverage, together with the procedure used to select the weight), as well as pseudocode for both the acquisition function and the hierarchical concept-graph construction. These additions will demonstrate that the reported gains arise from a principled combination rather than an arbitrary hyper-parameter choice. revision: yes

standing simulated objections not resolved

The proprietary thin-film defect dataset cannot be released, preventing full independent replication of the industrial-defect experiments that originally motivated the work.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents GSAL as a method that integrates a diffusion-based visual difficulty signal (reconstruction discrepancy and denoising variability) with a three-level concept graph for semantic coverage. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains are described that reduce the claimed gains in label efficiency or rare-class retrieval to the inputs by construction. The central claim is framed as the result of balancing two independent external signals, with no self-definitional steps or load-bearing internal reductions visible in the abstract or method outline.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities. The approach relies on standard diffusion models and concept graphs whose specific tuning or construction details are not given.

pith-pipeline@v0.9.0 · 5577 in / 1334 out tokens · 76772 ms · 2026-05-08T12:13:24.494459+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 5 canonical work pages

[1]

Deep batch active learning by diverse, uncertain gradient lower bounds

Jordan Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. InInternational Conference on Learning Representations (ICLR), 2019. 1, 3, 6, 7

2019
[2]

Diffusion active learn- ing: Towards data-driven experimental design in computed tomography.arXiv preprint arXiv:2504.03491, 2025

Luis Barba, Johannes Kirschner, Tomas Aidukas, Manuel Guizar-Sicairos, and Benjam´ın B´ejar. Diffusion active learn- ing: Towards data-driven experimental design in computed tomography.arXiv preprint arXiv:2504.03491, 2025. 3

work page arXiv 2025
[3]

The power of ensembles for active learning in image classification

William H Beluch, Tim Genewein, Andreas N ¨urnberger, and Jan M K¨orner. The power of ensembles for active learning in image classification. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9368–9377, 2018. 1, 2, 3

2018
[4]

Shedding light on large generative networks: Estimating epistemic uncer- tainty in diffusion models

Lucas Berry, Axel Brando, and David Meger. Shedding light on large generative networks: Estimating epistemic uncer- tainty in diffusion models. InUncertainty in Artificial Intel- ligence, pages 360–376. PMLR, 2024. 2, 3

2024
[5]

Neural-symbolic learning and reasoning: A survey and interpretation.arXiv preprint arXiv:1711.03902,

Tarek R Besold, Artur d’Avila Garcez, Sebastian Bader, Howard Bowman, Pedro Domingos, Pascal Hitzler, Kai- Uwe K ¨uhnberger, Luis C Lamb, Risto Miikkulainen, and Daniel L Silver. Neural-symbolic learning and reasoning: A survey and interpretation.arXiv preprint arXiv:1711.03902,

work page arXiv
[6]

Diffusion-based probabilistic un- certainty estimation for active domain adaptation.Advances in Neural Information Processing Systems, 36:17129–17155,

Zhekai Du and Jingjing Li. Diffusion-based probabilistic un- certainty estimation for active domain adaptation.Advances in Neural Information Processing Systems, 36:17129–17155,
[7]

The pascal visual object classes (voc) challenge.International Journal of Computer Vision (IJCV), 88(2):303–338, 2010

Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge.International Journal of Computer Vision (IJCV), 88(2):303–338, 2010. 6

2010
[8]

Deep bayesian active learning with image data

Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. InProceedings of the 34th International Conference on Machine Learning (ICML), pages 1183–1192, 2017. 1, 3, 6

2017
[9]

Neural-Symbolic Computing: An Effective Methodology for Principled Integration of Machine Learning and Reasoning

Artur d’Avila Garcez, Tarek R Besold, Luc De Raedt, Peter F ¨oldiak, Pascal Hitzler, Thomas Icard, Kai-Uwe K¨uhnberger, Luis C Lamb, Risto Miikkulainen, and Daniel L Silver.Neurosymbolic AI: The 3rd wave. arXiv preprint arXiv:1905.06088, 2019. 2, 3

work page Pith review arXiv 1905
[10]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. 6

2016
[11]

Active learn- ing for object detection with non-redundant informative sam- pling.arXiv preprint arXiv:2307.08414, 2023

Aral Hekimoglu, Adrian Brucker, Alper Kagan Kayali, Michael Schmidt, and Alvaro Marcos-Ramiro. Active learn- ing for object detection with non-redundant informative sam- pling.arXiv preprint arXiv:2307.08414, 2023. 1, 3, 6, 7

work page arXiv 2023
[12]

Segment any- thing

Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 4015–4026, 2023. 2

2023
[13]

Concept bottleneck models

Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. InInternational conference on machine learning, pages 5338–5348. PMLR, 2020. 3

2020
[14]

Diffusion-based deep active learning.arXiv preprint arXiv:2003.10339, 2020

Dan Kushnir and Luca Venturi. Diffusion-based deep active learning.arXiv preprint arXiv:2003.10339, 2020. 3

work page arXiv 2003
[15]

A sequential algorithm for training text classifiers

David D Lewis and William A Gale. A sequential algorithm for training text classifiers. InProceedings of the 17th An- nual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3–12, 1994. 1, 3

1994
[16]

Dong Liang, Jing-Wei Zhang, Ying-Peng Tang, and Sheng- Jun Huang. Mus-cdb: Mixed uncertainty sampling with class distribution balancing for active annotation in aerial object detection.IEEE Transactions on Geoscience and Remote Sensing, 61:1–13, 2023. 2, 3

2023
[17]

Microsoft coco: Common objects in context

Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEu- ropean Conference on Computer Vision (ECCV), pages 740– 755, 2014. 6

2014
[18]

Improved denoising diffusion probabilistic models

Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. InInternational conference on machine learning, pages 8162–8171. PMLR,
[19]

Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift.Advances in neural information processing systems, 32, 2019

Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshmi- narayanan, and Jasper Snoek. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift.Advances in neural information processing systems, 32, 2019. 2, 3

2019
[20]

Assemai: Interpretable image-based anomaly detection for manufacturing pipelines

Renjith Prasad, Chathurangi Shyalika, Fadi El Kalach, Re- vathy Venkataramanan, Ramtin Zand, Ramy Harik, and Amit Sheth. Assemai: Interpretable image-based anomaly detection for manufacturing pipelines. In2024 Interna- tional Conference on Machine Learning and Applications (ICMLA), pages 1720–1727. IEEE, 2024. 2

2024
[21]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the International Conference on Machine Learning (ICML),
[22]

Faster r-cnn: Towards real-time object detection with region proposal networks

Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. InAdvances in Neural Information Pro- cessing Systems (NeurIPS), 2015. 6

2015
[23]

Gradient-based quantification of epis- temic uncertainty for deep object detectors

Tobias Riedlinger, Matthias Rottmann, Marius Schubert, and Hanno Gottschalk. Gradient-based quantification of epis- temic uncertainty for deep object detectors. InProceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3921–3931, 2023. 2, 3

2023
[24]

High-resolution image syn- thesis with latent diffusion models

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 10684–10695, 2022. 2, 3, 5, 6

2022
[25]

Deep active learning for object detection

Soumya Roy, Asim Unmesh, and Vinay P Namboodiri. Deep active learning for object detection. InBMVC, page 375,
[26]

Overcoming data scarcity in biomedical imaging with a foundational multi- task model.Nature Computational Science, 4(7):495–509,

Raphael Sch ¨afer, Till Nicke, Henning H ¨ofener, Annkristin Lange, Dorit Merhof, Friedrich Feuerhake, V olkmar Schulz, Johannes Lotz, and Fabian Kiessling. Overcoming data scarcity in biomedical imaging with a foundational multi- task model.Nature Computational Science, 4(7):495–509,
[27]

Active learning for convolu- tional neural networks: A core-set approach

Ozan Sener and Silvio Savarese. Active learning for convolu- tional neural networks: A core-set approach. InInternational Conference on Learning Representations (ICLR), 2018. 1, 3, 6, 7

2018
[28]

Neurosymbolic ai-why, what, and how.IEEE Intelligent Systems, 38, 2023

Amit Sheth, Kaushik Roy, and Manas Gaur. Neurosymbolic ai-why, what, and how.IEEE Intelligent Systems, 38, 2023. 2, 3

2023
[29]

A comprehensive survey on rare event prediction

Chathurangi Shyalika, Ruwan Wickramarachchi, and Amit P Sheth. A comprehensive survey on rare event prediction. ACM Computing Surveys, 57(3):1–39, 2024. 2

2024
[30]

Entropy-based ac- tive learning for object detection with progressive diversity constraint

Jiaxi Wu, Jiaxin Chen, and Di Huang. Entropy-based ac- tive learning for object detection with progressive diversity constraint. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9397–9406,
[31]

Plug and Play Active Learning for Object Detection

Yang, Chenhongyi and Huang, Lichao and Crowley, Elliot J. Plug and Play Active Learning for Object Detection. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3, 6, 7

2024
[32]

Multiple instance active learning for object detection

Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5326–5335. IEEE, 2021. 1, 3, 6, 7

2021