High demand for computational power significantly limits the possibility of using modern deep learning methods in the environments where one has to deal with devices limited by the performance and the energy constraints. To address this issue, this paper proposes a novel method of combining the pruning and the transfer learning techniques for the purpose of delivering solid accuracy while simultaneously lowering the demand for energy and computing power. This method is referred to as \emph{Finicky Transfer Learning} (FTL) as it is finicky during the process of selecting filters from a pre-trained feature extractor to compose a sparser architecture. The proposed filter selection process is based on an original approach utilizing the Jaccard similarity coefficient calculated between the activation maps and the masks obtained by semantic segmentation. This enables the use of CNN, trained previously on a large generic dataset, in a crack classification task. The presented method significantly lowers the inference time while maintaining or even slightly increasing the classification accuracy, enabling real-time operation on single-board computers.
00350nas a2200133 4500000000100000000000100001008004100002260001100043100002000054700002000074245010000094490000700194022001500201 2021 d c6/2021 1 aBartosz Wójcik 1 aMateusz Żarski 00 aThe measurements of surface defect area with an RGB-D camera for a BIM-backed bridge inspection 0 v69 a0239-7528 01433nas a2200157 4500000000100000000000100001008004100002260001200043100002000055700002000075700002200095245009100117856003700208490000700245520102300252 2021 d c12/2021 1 aMateusz Żarski 1 aBartosz Wójcik 1 aJaroslaw Miszczak 00 aKrakN: Transfer Learning framework and dataset for infrastructure thin crack detection uhttps://arxiv.org/abs/2004.12337 0 v16 3 aMonitoring the technical condition of infrastructure is a crucial element of its maintenance. Although there are many deep learning models intended for this purpose, they are severely limited in their application due to labour-intensive gathering of new datasets and high demand for computing power during model training. To overcome these limiting factors we propose a KrakN framework. It enables end-to-end development of unique infrastructure defect detectors on digital images, achieving an accuracy of above 90%. The framework also supports the semi-automatic creation of new datasets and has modest computing power requirements. It can be used to immediately implement deep learning in the process of infrastructure management and, due to its architecture based on transfer learning, allows low metric loss when used across different datasets. We also demonstrate that thanks to its scalability and modular structure, the presented framework is easily modifiable, and can be used in realistic scenarios.
00551nas a2200169 4500000000100000000000100001008004100002260009400043100002000137700002000157700002000177700001800197700002200215245006500237856005700302020002200359 2021 d bInstitute of Fundamental Technological Research of the Polish Academy of Sciences aWarsaw 1 aMateusz Żarski 1 aBartosz Wójcik 1 aKamil Książek 1 aMarek Salamak 1 aJaroslaw Miszczak 00 aAdvances in applicable deep-learning based defect detection uhttp://v4shm.ippt.pan.pl/WEO2021-BookOfAbstracts.pdf a987-83-65550-23-1 00509nas a2200157 4500000000100000000000100001008004100002260009400043100002000137700002000157700001800177700002200195245005500217856005700272020002200329 2021 d bInstitute of Fundamental Technological Research of the Polish Academy of Sciences aWarsaw 1 aBartosz Wójcik 1 aMateusz Żarski 1 aMarek Salamak 1 aJaroslaw Miszczak 00 aExtracting crack characteristics from RGB-D images uhttp://v4shm.ippt.pan.pl/WEO2021-BookOfAbstracts.pdf a987-83-65550-23-1 01406nas a2200181 4500000000100000000000100001008004100002260001200043100002000055700002000075700002200095700002600117700002200143245008300165856003700248490000700285520093200292 2022 d c10/2022 1 aMateusz Żarski 1 aBartosz Wójcik 1 aJaroslaw Miszczak 1 aBartlomiej Blachowski 1 aMariusz Ostrowski 00 aComputer Vision Based Inspection on Post-Earthquake With UAV Synthetic Dataset uhttps://arxiv.org/abs/2210.05282 0 v10 3 aIn recent years, a lot of attention is paid to deep learning methods in the context of vision-based construction site safety systems. However, there is still more to be done to establish the relationship between supervised construction workers and their essential personal protective equipment, like hard hats. A deep learning method combining object detection, head center localization, and simple rule-based reasoning is proposed in this article. In tests, this solution surpassed the previous methods based on the relative bounding box position of different instances and direct detection of hard hat wearers and non-wearers. Achieving MS COCO style overall AP of 67.5% compared to 66.4% and 66.3% achieved by the approaches mentioned above, with class-specific AP for hard hat non-wearers of 64.1% compared to 63.0% and 60.3%. The results show that using deep learning methods with a humanly interpretable rule-based algorithm is better suited for detecting hard hat non-wearers.
01407nas a2200205 4500000000100000000000100001008004100002260002100043100001900064700002400083700002400107700002000131700002000151700002300171700002000194700001800214245007500232856004600307520084800353 2023 d aMontreal, Canada 1 aSamuel Kessler 1 aMateusz Ostaszewski 1 aMichał Bortkiewicz 1 aMateusz Żarski 1 aMaciej Wołczyk 1 aJack Parker-Holder 1 aStephen Roberts 1 aPiotr Miłoś 00 aThe Effectiveness of World Models for Continual Reinforcement Learning uhttps://lifelong-ml.cc/online_proceedings 3 aWorld models power some of the most efficient reinforcement learning algorithms. In this work, we showcase that they can be harnessed for continual learning – a situation when the agent faces changing environments. World models typically employ a replay buffer for training, which can be naturally extended to continual learning. We systematically study how different selective experience replay methods affect performance, forgetting, and transfer. We also provide recommendations regarding various modeling options for using world models. The best set of choices is called Continual-Dreamer, it is task-agnostic and utilizes the world model for continual exploration. Continual-Dreamer is sample efficient and outperforms state-of-the-art task-agnostic continual reinforcement learning methods on Minigrid and Minihack benchmarks.
01259nas a2200181 4500000000100000000000100001008004100002260002700043100002600070700002200096700002000118700002000138700002000158700002100178245011400199520074200313020002201055 2022 d c06/2022 bSpringer Link 1 aBartlomiej Blachowski 1 aMariusz Ostrowski 1 aMateusz Żarski 1 aBartosz Wójcik 1 aPiotr Tauzowski 1 aLukasz Jankowski 00 aAn Efficient Computer Vision-Based Method for Estimation of Dynamic Displacements in Spatial Truss Structures 3 aIn the present study a comparison of frequently used computer vision (CV)-based methods for structural health monitoring of truss structures is shown. The attention is paid to template matching methods that can be classified into one of two groups: area-based and feature-based methods. Synthetic but realistic video is used in this study. Results of the comparison are reliable due to the fact that the exact displacements are known from the finite element model of the investigated structure. From the variety of tested CV methods, the Kanade–Lucas–Tomasi algorithm with FREAK-based repetitive correction outperforms the remaining tested methods in terms of the computation time with a negligibly greater estimation error.
a978-3-031-07257-4 00403nas a2200157 4500000000100000000000100001008004100002260001200043100001800055700002000073700002600093700002100119245004800140856005000188490000700238 2025 d c02/2025 1 aDariusz Pojda 1 aMateusz Żarski 1 aAgnieszka Anna Tomaka 1 aLeszek Luchowski 00 adpVision: environment for multimodal images uhttp://dx.doi.org/10.1016/j.softx.2025.102093 0 v30 01549nas a2200121 4500000000100000000000100001008004100002260001200043100002000055700002200075245008900097520124100186 2024 d c09/2024 1 aMateusz Żarski 1 aJaroslaw Miszczak 00 aMulti-Step Feature Fusion for Natural Disaster Damage Assessment on Satellite Images 3 aQuick and accurate assessment of the damage state of buildings after natural disasters is crucial for undertaking properly targeted rescue and subsequent recovery operations, which can have a major impact on the safety of victims and the cost of disaster recovery. The quality of such a process can be significantly improved by harnessing the potential of machine learning methods in computer vision. This paper presents a novel damage assessment method using an original multi-step feature fusion network for the classification of the damage state of buildings based on pre- and post-disaster large-scale satellite images. We introduce a novel convolutional neural network (CNN) module that performs feature fusion at multiple network levels between pre- and post-disaster images in the horizontal and vertical directions of CNN network. An additional network element – Fuse Module – was proposed to adapt any CNN model to analyze image pairs in the issue of pair classification. We use, open, large-scale datasets (IDA-BD and xView2) to verify, that the proposed method is suitable to improve on existing state-of-the-art architectures. We report over a 3 percentage point increase in the accuracy of the Vision Transformer model. 00357nas a2200121 4500000000100000000000100001008004100002100002000043700002400063700002000087245010700107856002100214 d 1 aKatarzyna Filus 1 aMichał Romaszewski 1 aMateusz Żarski 00 aSemantic Depth Matters: Explaining Errors of Deep Vision Networks through Perceived Class Similarities uarXiv:2504.09956 01470nas a2200145 4500000000100000000000100001008004100002260004800043100002000091700002300111245007400134856004900208520104500257020002201302 2025 d c05/2025 bSpringer Nature aKonstanz, Germany 1 aMateusz Żarski 1 aSławomir Nowaczyk 00 aBalancing Performance and Scalability of Demand Forecasting ML Models uhttps://doi.org/10.1007/978-3-031-91398-3_10 3 aBalancing performance and scalability is a major concern when developing robust ML models for diverse, big-data scenarios, such as predicting demand for a number of products across multiple locations. The two mutually opposite approaches are to use a single ML model for maximizing scalability, often at the expense of performance, or to use a specialized model for each specific use case, which is often prohibitive in terms of computational costs. In this paper, we propose to balance those two approaches using our methods of model clustering and grouping. We achieve the performance level of a single use-case model while preserving the global scalability of the solution. In our experiments, we use a publicly available demand forecasting dataset as a use case. We develop and train baseline shallow ML models and DL models for both maximizing performance and scalability. Then, we showcase a desirable balance that can be achieved using our proposed methods, one that outperforms both shallow ML and specific use-case models.
a978-3-031-91398-3 01423nas a2200121 4500000000100000000000100001008004100002100002000043700002000063245012200083856003700205520105900242 d 1 aKatarzyna Filus 1 aMateusz Żarski 00 aBeyond Accuracy: Uncovering the Role of Similarity Perception and its Alignment with Semantics in Supervised Learning uhttps://arxiv.org/abs/2505.21338 3 aSimilarity manifests in various forms, including semantic similarity that is particularly important, serving as an approximation of human object categorization based on e.g. shared functionalities and evolutionary traits. It also offers practical advantages in computational modeling via lexical structures such as WordNet with constant and interpretable similarity. As in the domain of deep vision, there is still not enough focus on the phenomena regarding the similarity perception emergence. We introduce Deep Similarity Inspector (DSI) -- a systematic framework to inspect how deep vision networks develop their similarity perception and its alignment with semantic similarity. Our experiments show that both Convolutional Neural Networks' (CNNs) and Vision Transformers' (ViTs) develop a rich similarity perception during training with 3 phases (initial similarity surge, refinement, stabilization), with clear differences between CNNs and ViTs. Besides the gradual mistakes elimination, the mistakes refinement phenomenon can be observed.
01497nas a2200133 4500000000100000000000100001008004100002260001200043100002000055700002300075245007600098856005900174520113000233 2026 d c07/2026 1 aMateusz Żarski 1 aSławomir Nowaczyk 00 aNeuroplasticity-inspired dynamic ANNs for multi-task demand forecasting uhttps://ieeexplore.ieee.org/abstract/document/11592373 3 aThis paper introduces a novel approach to Dynamic Artificial Neural Networks (D-ANNs) for multi-task demand forecasting called Neuroplastic Multi-Task Network (NMT-Net). Unlike conventional methods focusing on inference-time dynamics or computational efficiency, our proposed method enables structural adaptability of the computational graph during training, inspired by neuroplasticity as seen in biological systems. Each new task triggers a dynamic network adaptation, including similarity-based task identification and selective training of candidate ANN heads, which are then assessed and integrated into the model based on their performance. We evaluated our framework using three real-world multi-task demand forecasting datasets from Kaggle. We demonstrated its superior performance and consistency, achieving lower RMSE and standard deviation compared to traditional baselines and state-of-the-art multi-task learning methods. NMT-Net offers a scalable, adaptable solution for multi-task and continual learning in time series prediction. The complete code for NMT-Net is available from our GitHub repository.
01927nas a2200181 4500000000100000000000100001008004100002260001200043100002200055700002600077700002000103700002000123700002000143700002200163245011400185856006400299520138200363 2023 d c01/2023 1 aMariusz Ostrowski 1 aBartlomiej Blachowski 1 aBartosz Wójcik 1 aMateusz Żarski 1 aPiotr Tauzowski 1 aŁukasz Jankowski 00 aA framework for computer vision-based health monitoring of a truss structure subjected to unknown excitations uhttps://link.springer.com/article/10.1007/s11803-023-2154-3 3 aComputer vision (CV) methods for measurement of structural vibration are less expensive, and their application is more straightforward than methods based on sensors that measure physical quantities at particular points of a structure. However, CV methods produce significantly more measurement errors. Thus, computer vision-based structural health monitoring (CVSHM) requires appropriate methods of damage assessment that are robust with respect to highly contaminated measurement data. In this paper a complete CVSHM framework is proposed, and three damage assessment methods are tested. The first is the augmented inverse estimate (AIE), proposed by Peng et al. in 2021. This method is designed to work with highly contaminated measurement data, but it fails with a large noise provided by CV measurement. The second method, as proposed in this paper, is based on the AIE, but it introduces a weighting matrix that enhances the conditioning of the problem. The third method, also proposed in this paper, introduces additional constraints in the optimization process; these constraints ensure that the stiff ness of structural elements can only decrease. Both proposed methods perform better than the original AIE. The latter of the two proposed methods gives the best results, and it is robust with respect to the selected coefficients, as required by the algorithm.
01850nas a2200121 4500000000100000000000100001008004100002100001800043700002300061700002000084245008900104520153500193 2025 d 1 aAnna Zawadzka 1 aPrzemysław Głomb 1 aMateusz Żarski 00 aDevelopment of an effective method for detecting small buildings in satellite images 3 aDetecting and monitoring small-scale buildings is a critical challenge in various sectors and can be crucial for various important tasks. Buildings considered small are those with a footprint smaller than 10 × 10 meters, a definition justified by the resolution of Sentinel-2 images, which is 10 meters at best. Examples of tasks for which small building detection is essential include assessment and monitoring of damage caused by natural disasters or war, identification of illegal settlements, detecting new buildings in key military areas, and identification of structures in protected areas. Regrettably, widely available satellite imagery, such as Sentinel-2, provides very limited spatial resolution, making it impossible to efficiently detect small structures. Despite its importance, this issue remains unexplored, with very limited research directly concerning building segmentation and detecting small structures. Traditional segmentation and classification methods exhibit severe limitations, particularly due to small buildings frequently being positioned at pixel boundaries, essentially occupying an insignificant fraction of the original pixel. Additionally, in the context of buildings in complex environments (e.g., surrounded by vegetation or shadows), popular methods such as U-Net with ResNet often fail to deliver reliable results. In this study, we propose a novel approach to this issue by integrating a state-of-the-art object detection model with a deep-learning-based super-resolution network.
00365nas a2200145 4500000000100000000000100001008004100002100001800043700002000061700001700081700002300098700002400121700002000145245005400165 2026 d 1 aAnna Zawadzka 1 aMateusz Żarski 1 aKamil Drejer 1 aPrzemysław Głomb 1 aMichał Romaszewski 1 aMichał Cholewa 00 a Map-Guided Cross-Training for Building Detection 00258nas a2200085 4500000000100000008004100001100002000042700002000062245009000082 2026 d 1 aKatarzyna Filus 1 aMateusz Żarski 00 aInspecting Training Dynamics of Similarity Development in Supervised Vision Networks