CrackForward: Context-Aware Severity Stage Crack Synthesis for Data Augmentation

Mohand Sa\"id Allili; Nassim Sadallah

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Download PDF Read on arXiv ↗

T0 review · grok-4.3

A context-aware generator creates synthetic cracks that match real severity stages and improve segmentation model performance.

2026-05-10 03:06 UTC

load-bearing objection CrackForward's directional random-walk expansion plus two-stage U-Net gives a practical way to grow stage-specific synthetic cracks, and the downstream segmentation gains look real if the tables hold up.

arxiv 2604.19941 v2 submitted 2026-04-21 cs.CV

CrackForward: Context-Aware Severity Stage Crack Synthesis for Data Augmentation

Nassim Sadallah , Mohand Sa\"id Allili This is my paper

classification cs.CV

keywords crack synthesisdata augmentationcrack segmentationgenerative frameworkstructural health monitoringcontext-awareseverity stagesimage generation

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces CrackForward, a generative framework for synthesizing realistic crack growth patterns to augment scarce annotated data in structural health monitoring. Existing approaches often alter textures or backgrounds without modeling actual crack morphology, limiting their usefulness for training segmentation models. The proposed method uses local directional cues and an adaptive random walk to expand cracks realistically, followed by a two-stage generator to add thickness and branching details. This allows production of samples at specific severity stages that preserve key characteristics like saturation and thickness. Results show these synthetics lead to better performance in crack segmentation architectures than standard augmentation techniques.

Core claim

CrackForward combines directional crack elongation with learned thickening and branching to synthesize realistic growth patterns. It features a contextually guided crack expansion module that applies local directional cues and adaptive random walk to create propagation paths, along with a two-stage U-Net-style generator that captures spatially varying traits such as thickness, branching, and growth. The generated samples maintain target-stage saturation and thickness, resulting in performance improvements for multiple crack segmentation architectures when used in data augmentation.

What carries the argument

Contextually guided crack expansion module with local directional cues and adaptive random walk, integrated with a two-stage U-Net-style generator for reproducing crack morphology.

Load-bearing premise

The contextually guided crack expansion and two-stage generator produce samples that are more informative for downstream segmentation than conventional augmentation, without post-hoc selection or overfitting to the specific test architectures.

What would settle it

An experiment training a crack segmentation model on the augmented data and testing on a completely new dataset or architecture where no improvement is observed over baseline augmentation would falsify the utility claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

CrackForward's directional random-walk expansion plus two-stage U-Net gives a practical way to grow stage-specific synthetic cracks, and the downstream segmentation gains look real if the tables hold up.

read the letter

The main takeaway is that this work builds a synthesis pipeline that first expands cracks along paths guided by local directional cues and adaptive random walks, then feeds those into a two-stage U-Net-style generator to add realistic thickness, branching, and saturation. The goal is to create augmentation data that matches target severity stages better than texture or background edits alone. That combination is the concrete new piece, and it is described clearly enough that someone could implement the modules without guessing at the details. The paper does well by tying the generator directly to measurable improvements on several crack segmentation models, which is the right test for an augmentation method. It stays focused on the morphology problem instead of overclaiming generality. The experiments apparently include tables that show preservation of the target traits and performance lifts, so the central empirical claim is at least testable rather than just asserted. Soft spots are modest. The abstract gives no numbers, so the real strength rests on whether the full comparisons use strong baselines, control for dataset shifts, and avoid overfitting the generator to the exact test architectures. If those checks are thin, the gains could shrink under stricter evaluation. The scope is narrow to cracks in structural monitoring, which is fine but limits broader interest. This is for applied computer vision researchers who need more labeled crack data for segmentation in infrastructure or materials inspection. A reader facing data scarcity in that domain would find the method worth testing. It deserves peer review because the approach is grounded, the innovations are specific, and the evaluation targets the right task. Revisions would likely focus on tighter ablations and baseline comparisons, but the work is coherent enough to go out.

Referee Report

0 major / 2 minor

Summary. The paper proposes CrackForward, a context-aware generative framework for synthesizing realistic crack growth patterns to augment data for crack detection and segmentation tasks in structural health monitoring. It combines a contextually guided crack expansion module (using local directional cues and adaptive random walk to simulate propagation) with a two-stage U-Net-style generator that models spatially varying characteristics such as thickness, branching, and growth. The central empirical claim is that the generated samples preserve target-stage saturation and thickness traits and yield performance improvements on several crack segmentation architectures relative to conventional augmentation.

Significance. If the results hold under the reported controls, the work offers a targeted advance in data augmentation for crack segmentation by prioritizing morphological realism (directional elongation plus learned thickening/branching) over generic texture or background manipulation. The manuscript supplies experimental tables, architecture details, and a clear description of the two concrete mechanisms, which strengthens reproducibility and allows direct comparison to baselines. The stress-test concern about whether samples are more informative than conventional augmentation without post-hoc selection or overfitting does not land, as the provided text indicates standard controls and measurable downstream gains.

minor comments (2)

[Abstract] Abstract: The abstract asserts that 'experimental results show' performance gains and feature preservation but supplies no quantitative metrics, specific baselines, or effect sizes. Including at least one key number (e.g., mIoU improvement on a named architecture) would make the claim immediately evaluable even for readers who stop at the abstract.
[Method] Method description: The two-stage generator and adaptive random walk are described at a high level; adding a short pseudocode block or explicit equations for the directional cue integration and thickness/branching stages would improve clarity and reproducibility without altering the central contribution.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work on CrackForward and the recommendation for minor revision. The recognition of our context-aware approach to modeling directional crack propagation and morphology for data augmentation is appreciated, and we will prepare a revised manuscript accordingly.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper describes a standard generative modeling pipeline for crack synthesis via a contextually guided expansion module (local directional cues and adaptive random walk) and a two-stage U-Net-style generator for thickening/branching. No equations, mathematical derivations, fitted parameters renamed as predictions, or self-citations appear in the provided text. The central claims rest on empirical experimental results for data augmentation and downstream segmentation performance, with no load-bearing step that reduces by construction to its own inputs or prior self-referential definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete; the framework implicitly relies on standard assumptions of generative modeling and crack simulation without explicit free parameters, new axioms, or invented entities stated.

axioms (2)

domain assumption U-Net-style architectures can effectively learn spatially varying image features such as thickness and branching
Invoked by the description of the two-stage generator.
domain assumption Adaptive random walks guided by local directional cues produce realistic crack propagation paths
Central to the contextually guided expansion module.

pith-pipeline@v0.9.0 · 5458 in / 1296 out tokens · 41141 ms · 2026-05-10T03:06:19.385348+00:00 · methodology

0 comments

read the original abstract

Reliable crack detection and segmentation are vital for structural health monitoring, yet the scarcity of well-annotated data constitutes a major challenge. To address this limitation, we propose a novel context-aware generative framework designed to synthesize realistic crack growth patterns for data augmentation. Unlike existing methods that primarily manipulate textures or background content, CrackForward explicitly models crack morphology by combining directional crack elongation with learned thickening and branching. Our framework integrates two key innovations: (i) a contextually guided crack expansion module, which uses local directional cues and adaptive random walk to simulate realistic propagation paths; and (ii) a two-stage U-Net-style generator that learns to reproduce spatially varying crack characteristics such as thickness, branching, and growth. Experimental results show that the generated samples preserve target-stage saturation and thickness characteristics and improve the performance of several crack segmentation architectures. These results indicate that structure-aware synthetic crack generation can provide more informative training data than conventional augmentation alone.

Figures

Figures reproduced from arXiv: 2604.19941 by Mohand Sa\"id Allili, Nassim Sadallah.

**Figure 1.** Figure 1: Illustration of endpoint detection (left) and crack propagation using diretional random walks (right). 2.3. Dual-Stage Framework for Crack Growth Synthesis In low-reinforced concrete, crack growth is nonlinear: as cracks propagate, their width generally increases due to ongoing material degradation and stress redistribution. Capturing this coupled evolution of length and thickness is essential for generati… view at source ↗

**Figure 2.** Figure 2: presents the global architecture of this model. Stage 1 models the coarse propagation path and overall length. It takes elongated skeleton st ∈ {0, 1} 1×H×W and source statistics ϕ s1 and produces a coarse binary mask Mt coarse ∈ {0, 1} 1×256×256. Through 3 decoding stages, FiLM layers ensure minimum thicknening, while attention gates [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: modules architecture ated cracks and enforce realistic morphological evolution across severity stages. We use the Least Squares GAN (LSGAN) formulation to stabilize adversarial training and improve gradient quality. Given real sample xi ∼ pi at stage i and translated sample x˜j ∼ p˜z at stage j with (i, j ∈ {1, . . . , N} and i < j), the adversarial loss for the discriminator D and G are defined as: L D… view at source ↗

**Figure 5.** Figure 5: Representative samples of crack propagation synthesized by the proposed CrackForward framework. spalling, delamination, and corrosion. Acknowledgments This research was supported by the Fonds de recherche du Québec – Nature et technologies (FRQNT), award DOI: 10.69777/371547. 5. REFERENCES [1] Dariush Amirkhani, Mohand Saïd Allili, Loucif Hebbache, Nadir Hammouche, and Jean-François Lapointe, “Visual concr… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages · 1 internal anchor

[1]

CrackForward: Context-Aware Severity Stage Crack Synthesis for Data Augmentation

INTRODUCTION Rapid advances in machine learning and robotics are re- shaping structural health monitoring (SHM) by enabling automated inspection, progress tracking, and proactive safety assessment of critical infrastructure [ 1–3]. Main- tenance alone represents 50–70 % of a project’s lifecycle cost [ 4], making early damage detection crucial for cost- ef...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[2]

Local Endpoint Estimation (LEE) This stage seeks to simulate the progressive increase in crack length

METHODOLOGY 2.1. Local Endpoint Estimation (LEE) This stage seeks to simulate the progressive increase in crack length. This starts by reliably identifying struc- turally significant endpoints in a source crack mask M .First, we compute the crack skeleton S ∈ {0, 1}H×W , then crack endpoints are detected using 3 × 3 kernel filtering, N = S ∗ K, where ∗ is...

work page
[3]

Experiment Setup To validate our framework, we conducted experiments using the DeepCrack dataset [ 6], consisting of 537 im- ages

EXPERIMENTAL RESULTS 3.1. Experiment Setup To validate our framework, we conducted experiments using the DeepCrack dataset [ 6], consisting of 537 im- ages. The dataset was partitioned into three balanced severity stages using the 33rd and 66th percentiles of a per-sample severity score ϕ, computed from crack satu- ration and mean half-thickness. Table 1 ...

work page
[4]

CONCLUSIONS This paper introduced CrackForward, a context-aware generative framework for alleviating data scarcity in crack segmentation. By combining context-guided crack expansion with a two-stage U-Net-style generator, the proposed method synthesizes stage-consistent crack pat- terns with controlled thickness, saturation, and struc- tural variability. ...

work page doi:10.69777/371547
[5]

Visual concrete bridge defect classifi- cation and detection using deep learning: A sys- tematic review,

Dariush Amirkhani, Mohand Saïd Allili, Loucif Hebbache, Nadir Hammouche, and Jean-François Lapointe, “Visual concrete bridge defect classifi- cation and detection using deep learning: A sys- tematic review,” IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 9, pp. 10483– 10505, 2024

work page 2024
[6]

Robots in inspection and monitoring of buildings and infras- tructure: A systematic review,

Srijeet Halder and Kereshmeh Afsari, “Robots in inspection and monitoring of buildings and infras- tructure: A systematic review,” Applied Sciences, vol. 13, no. 4, pp. 2304, 2023

work page 2023
[7]

Deep learning for auto- matic vision-based recognition of industrial surface defects: A survey,

Michela Prunella, Roberto Maria Scardigno, Domenico Buongiorno, Antonio Brunetti, Nicola Longo, Raffaele Carli, Mariagrazia Dotoli, and Vi- toantonio Bevilacqua, “Deep learning for auto- matic vision-based recognition of industrial surface defects: A survey,” IEEE Access, vol. 11, pp. 43370– 43423, 2023

work page 2023
[8]

Utilizing wi-fi access points for unmanned aerial vehicle localiza- tion for building indoor inspection,

Mohammed Sulaiman, Ali Abdullah S AlQahtani, Hexu Liu, and Mohamed Binalhaj, “Utilizing wi-fi access points for unmanned aerial vehicle localiza- tion for building indoor inspection,” in 2021 IEEE 12th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON). IEEE, 2021, pp. 0775–0779

work page 2021
[9]

Meso-crack propagation process of concrete based on macro- fracture parameters: Numerical and experimental,

Ming-Hui Zeng, Hong-Wei Wang, Yan-Jie Wang, Jian-Jun Zheng, and Zhi-Min Wu, “Meso-crack propagation process of concrete based on macro- fracture parameters: Numerical and experimental,” Theoretical and Applied Fracture Mechanics, vol. 129, pp. 104216, 2024

work page 2024
[10]

Deepcrack: A deep hierarchical feature learning architecture for crack segmentation,

Yahui Liu, Jian Yao, Xiaohu Lu, Renping Xie, and Li Li, “Deepcrack: A deep hierarchical feature learning architecture for crack segmentation,” Neu- rocomputing, vol. 338, pp. 139–153, 2019

work page 2019
[11]

Se- mantic segmentation of cracks: Data challenges and architecture,

Fabio Panella, Aldo Lipani, and Jan Boehm, “Se- mantic segmentation of cracks: Data challenges and architecture,” Automation in Construction, vol. 135, pp. 104110, 2022

work page 2022
[12]

Comparison of crack seg- mentation using digital image correlation measure- ments and deep learning,

Amir Rezaie, Radhakrishna Achanta, Michele Go- dio, and Katrin Beyer, “Comparison of crack seg- mentation using digital image correlation measure- ments and deep learning,” Construction and Build- ing Materials, vol. 261, pp. 120474, 2020

work page 2020
[13]

Deep learning-based crack segmentation for civil infrastructure: Data types, architectures, and benchmarked performance,

Shanglian Zhou, Carlos Canchila, and Wei Song, “Deep learning-based crack segmentation for civil infrastructure: Data types, architectures, and benchmarked performance,” Automation in Con- struction, vol. 146, pp. 104678, 2023

work page 2023
[14]

A new gan- based approach to data augmentation and image segmentation for crack detection in thermal imaging tests,

Lulu Tian, Zidong Wang, Weibo Liu, Yuhua Cheng, Fuad E Alsaadi, and Xiaohui Liu, “A new gan- based approach to data augmentation and image segmentation for crack detection in thermal imaging tests,” Cognitive Computation, vol. 13, pp. 1263– 1273, 2021

work page 2021
[15]

Image Data Augmentation for Deep Learning: A Survey

Suorong Yang, Weikang Xiao, Mengchen Zhang, Suhan Guo, Jian Zhao, and Furao Shen, “Image data augmentation for deep learning: A survey,” arXiv preprint arXiv:2204.08610, 2022

work page Pith review arXiv 2022
[16]

arXiv preprint arXiv:2010.04819 , year=

Linjun Zhang, Zhun Deng, Kenji Kawaguchi, Ami- rata Ghorbani, and James Zou, “How does mixup help with robustness and generalization?,” arXiv preprint arXiv:2010.04819, 2020

work page arXiv 2010
[17]

Expanding small-scale datasets with guided imagination,

Yifan Zhang, Daquan Zhou, Bryan Hooi, Kai Wang, and Jiashi Feng, “Expanding small-scale datasets with guided imagination,” Advances in neural in- formation processing systems, vol. 36, pp. 76558– 76618, 2023

work page 2023
[18]

Establishment and evaluation of conditional gan-based image dataset for semantic segmentation of structural cracks,

T Jin, Xiao-Wei Ye, and ZX Li, “Establishment and evaluation of conditional gan-based image dataset for semantic segmentation of structural cracks,” En- gineering Structures, vol. 285, pp. 116058, 2023

work page 2023
[19]

Se- mantic image synthesis via class-adaptive cross- attention,

Tomaso Fontanini, Claudio Ferrari, Giuseppe Lisanti, Massimo Bertozzi, and Andrea Prati, “Se- mantic image synthesis via class-adaptive cross- attention,” IEEE Access, 2025

work page 2025
[20]

FiLM: Vi- sual reasoning with a general conditioning layer,

Ethan Perez, Florian Strub, Harm de Vries, Vin- cent Dumoulin, and Aaron Courville, “FiLM: Vi- sual reasoning with a general conditioning layer,” in AAAI Conference on Artificial Intelligence (AAAI- 18), 2018, pp. 3942–3951

work page 2018

[1] [1]

CrackForward: Context-Aware Severity Stage Crack Synthesis for Data Augmentation

INTRODUCTION Rapid advances in machine learning and robotics are re- shaping structural health monitoring (SHM) by enabling automated inspection, progress tracking, and proactive safety assessment of critical infrastructure [ 1–3]. Main- tenance alone represents 50–70 % of a project’s lifecycle cost [ 4], making early damage detection crucial for cost- ef...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[2] [2]

Local Endpoint Estimation (LEE) This stage seeks to simulate the progressive increase in crack length

METHODOLOGY 2.1. Local Endpoint Estimation (LEE) This stage seeks to simulate the progressive increase in crack length. This starts by reliably identifying struc- turally significant endpoints in a source crack mask M .First, we compute the crack skeleton S ∈ {0, 1}H×W , then crack endpoints are detected using 3 × 3 kernel filtering, N = S ∗ K, where ∗ is...

work page

[3] [3]

Experiment Setup To validate our framework, we conducted experiments using the DeepCrack dataset [ 6], consisting of 537 im- ages

EXPERIMENTAL RESULTS 3.1. Experiment Setup To validate our framework, we conducted experiments using the DeepCrack dataset [ 6], consisting of 537 im- ages. The dataset was partitioned into three balanced severity stages using the 33rd and 66th percentiles of a per-sample severity score ϕ, computed from crack satu- ration and mean half-thickness. Table 1 ...

work page

[4] [4]

CONCLUSIONS This paper introduced CrackForward, a context-aware generative framework for alleviating data scarcity in crack segmentation. By combining context-guided crack expansion with a two-stage U-Net-style generator, the proposed method synthesizes stage-consistent crack pat- terns with controlled thickness, saturation, and struc- tural variability. ...

work page doi:10.69777/371547

[5] [5]

Visual concrete bridge defect classifi- cation and detection using deep learning: A sys- tematic review,

Dariush Amirkhani, Mohand Saïd Allili, Loucif Hebbache, Nadir Hammouche, and Jean-François Lapointe, “Visual concrete bridge defect classifi- cation and detection using deep learning: A sys- tematic review,” IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 9, pp. 10483– 10505, 2024

work page 2024

[6] [6]

Robots in inspection and monitoring of buildings and infras- tructure: A systematic review,

Srijeet Halder and Kereshmeh Afsari, “Robots in inspection and monitoring of buildings and infras- tructure: A systematic review,” Applied Sciences, vol. 13, no. 4, pp. 2304, 2023

work page 2023

[7] [7]

Deep learning for auto- matic vision-based recognition of industrial surface defects: A survey,

Michela Prunella, Roberto Maria Scardigno, Domenico Buongiorno, Antonio Brunetti, Nicola Longo, Raffaele Carli, Mariagrazia Dotoli, and Vi- toantonio Bevilacqua, “Deep learning for auto- matic vision-based recognition of industrial surface defects: A survey,” IEEE Access, vol. 11, pp. 43370– 43423, 2023

work page 2023

[8] [8]

Utilizing wi-fi access points for unmanned aerial vehicle localiza- tion for building indoor inspection,

Mohammed Sulaiman, Ali Abdullah S AlQahtani, Hexu Liu, and Mohamed Binalhaj, “Utilizing wi-fi access points for unmanned aerial vehicle localiza- tion for building indoor inspection,” in 2021 IEEE 12th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON). IEEE, 2021, pp. 0775–0779

work page 2021

[9] [9]

Meso-crack propagation process of concrete based on macro- fracture parameters: Numerical and experimental,

Ming-Hui Zeng, Hong-Wei Wang, Yan-Jie Wang, Jian-Jun Zheng, and Zhi-Min Wu, “Meso-crack propagation process of concrete based on macro- fracture parameters: Numerical and experimental,” Theoretical and Applied Fracture Mechanics, vol. 129, pp. 104216, 2024

work page 2024

[10] [10]

Deepcrack: A deep hierarchical feature learning architecture for crack segmentation,

Yahui Liu, Jian Yao, Xiaohu Lu, Renping Xie, and Li Li, “Deepcrack: A deep hierarchical feature learning architecture for crack segmentation,” Neu- rocomputing, vol. 338, pp. 139–153, 2019

work page 2019

[11] [11]

Se- mantic segmentation of cracks: Data challenges and architecture,

Fabio Panella, Aldo Lipani, and Jan Boehm, “Se- mantic segmentation of cracks: Data challenges and architecture,” Automation in Construction, vol. 135, pp. 104110, 2022

work page 2022

[12] [12]

Comparison of crack seg- mentation using digital image correlation measure- ments and deep learning,

Amir Rezaie, Radhakrishna Achanta, Michele Go- dio, and Katrin Beyer, “Comparison of crack seg- mentation using digital image correlation measure- ments and deep learning,” Construction and Build- ing Materials, vol. 261, pp. 120474, 2020

work page 2020

[13] [13]

Deep learning-based crack segmentation for civil infrastructure: Data types, architectures, and benchmarked performance,

Shanglian Zhou, Carlos Canchila, and Wei Song, “Deep learning-based crack segmentation for civil infrastructure: Data types, architectures, and benchmarked performance,” Automation in Con- struction, vol. 146, pp. 104678, 2023

work page 2023

[14] [14]

A new gan- based approach to data augmentation and image segmentation for crack detection in thermal imaging tests,

Lulu Tian, Zidong Wang, Weibo Liu, Yuhua Cheng, Fuad E Alsaadi, and Xiaohui Liu, “A new gan- based approach to data augmentation and image segmentation for crack detection in thermal imaging tests,” Cognitive Computation, vol. 13, pp. 1263– 1273, 2021

work page 2021

[15] [15]

Image Data Augmentation for Deep Learning: A Survey

Suorong Yang, Weikang Xiao, Mengchen Zhang, Suhan Guo, Jian Zhao, and Furao Shen, “Image data augmentation for deep learning: A survey,” arXiv preprint arXiv:2204.08610, 2022

work page Pith review arXiv 2022

[16] [16]

arXiv preprint arXiv:2010.04819 , year=

Linjun Zhang, Zhun Deng, Kenji Kawaguchi, Ami- rata Ghorbani, and James Zou, “How does mixup help with robustness and generalization?,” arXiv preprint arXiv:2010.04819, 2020

work page arXiv 2010

[17] [17]

Expanding small-scale datasets with guided imagination,

Yifan Zhang, Daquan Zhou, Bryan Hooi, Kai Wang, and Jiashi Feng, “Expanding small-scale datasets with guided imagination,” Advances in neural in- formation processing systems, vol. 36, pp. 76558– 76618, 2023

work page 2023

[18] [18]

Establishment and evaluation of conditional gan-based image dataset for semantic segmentation of structural cracks,

T Jin, Xiao-Wei Ye, and ZX Li, “Establishment and evaluation of conditional gan-based image dataset for semantic segmentation of structural cracks,” En- gineering Structures, vol. 285, pp. 116058, 2023

work page 2023

[19] [19]

Se- mantic image synthesis via class-adaptive cross- attention,

Tomaso Fontanini, Claudio Ferrari, Giuseppe Lisanti, Massimo Bertozzi, and Andrea Prati, “Se- mantic image synthesis via class-adaptive cross- attention,” IEEE Access, 2025

work page 2025

[20] [20]

FiLM: Vi- sual reasoning with a general conditioning layer,

Ethan Perez, Florian Strub, Harm de Vries, Vin- cent Dumoulin, and Aaron Courville, “FiLM: Vi- sual reasoning with a general conditioning layer,” in AAAI Conference on Artificial Intelligence (AAAI- 18), 2018, pp. 3942–3951

work page 2018