This study evaluates the performance of selected machine learning methods, Maximum Likelihood (MLC), Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Machine (SVM), and Artificial Neural Networks (ANN), for land use/land cover (LULC) classification using Sentinel-2 satellite imagery. Each algorithm was tested across multiple classification scenarios that systematically varied both training sample size and training sample quality. To emulate realistic imperfections in reference data, controlled levels of label noise were introduced into the training set, and the resulting changes in classification performance were analysed. In addition, each model’s susceptibility to overfitting was assessed by comparing performance on training data and independent test data. To reduce the influence of a single random draw of training and testing pixels, all sampling-based experiments were repeated 10 times using different random seed values, and the reported results were summarized using repeated-run statistics. This design enabled a comparable assessment of how classification accuracy, stability, and generalization depend on dataset size and quality. The results indicate that SVM was the most consistent and reliable method across the tested scenarios, achieving high classification accuracy over a wide range of training sample sizes and maintaining strong robustness under noisy labels. Compared with the other methods, SVM showed lower sensitivity to degraded training data and a smaller tendency to overfit, which makes it a strong baseline choice when reference data are limited or imperfect.
This study aims to couple satellite data with precise geometric models. The performance assessment of the method is based on the dataset collected at a part of the vertical base geodetic control network in the Podkarpackie province in Poland. We used baselines determined in two 12-hour-long static Global Navigation Satellite System (GNSS) measurement sessions and reference coordinates of vertical base points of the geodetic control network. Baselines obtained from satellite measurements were used for the fusion, considering the height anomalies taken from the corresponding official geoid model PL-geoid2021. Proprietary solutions played a key role in satellite measurements by ensuring the correct performance of long-term measurements without requiring the manual measurement of the height of GNSS antennas relative to the marks of the measured geodetic control points. Two patented solutions were used for the first time in this study. The results confirmed the correctness and high applicability of the presented approach.
Accurate indoor camera localization is crucial for applications in augmented reality, robotics, and autonomous navigation. While single-image deep learning models for 6-DOF pose regression have shown competitive results on established benchmarks, their development still requires extensive data annotation and hyperparameter tuning. In this work, we investigate the combination of advanced network architectures, transfer learning, and synthetic data to improve single-image indoor pose regression. Our approach employs a ResNet50 backbone pre-trained on the Places365 dataset and further trained and evaluated on established benchmarks. To enhance the training data, synthetic images are generated from 3D BIM models using Unreal Engine, with alignment procedures ensuring accurate correspondence between synthetic and real environments. Real RGB images are preprocessed to resemble synthetic data, enabling effective cross-domain evaluation. Experiments demonstrate that both architectural design and pretraining significantly influence model performance. On the UniMelb dataset (real-to-real scenario), the model achieves 0.21 m and 0.80 degrees errors, surpassing baseline accuracy. We also present cross-validation and synthetic-to-synthetic experiments, providing insights into factors affecting performance and interactions between architecture, pretraining, and dataset characteristics.
This study presents a comparative analysis of two three-dimensional reconstruction models of Powder Magazine no. 2 of the 19h-century Cracow Fortress, developed using terrestrial laser scanning (TLS) and close-range digital photogrammetry to assess the suitability of both methods for the documentation and reconstruction of heritage structures, as well as to evaluate their geometric accuracy and visual quality. The investigated structure was scanned from four positions using a Leica ScanStation P40 and photographed using a Canon EOS 400D camera. Based on the acquired point clouds, two 3D models were generated: a TLS-based model in Autodesk Revit and a photogrammetric model in Trimble SketchUp, both at Level of Detail (LOD) 200. Accuracy assessment was conducted in CloudCompare using the Cloud-to-Cloud (C2C) method, supplemented by visual and historical comparisons. The results show high geometric consistency between the datasets (with deviations below 0.2 m) and confirm the effectiveness and complementarity of both techniques in the documentation of cultural heritage. The developed models provide valuable material for scientific research, conservation planning, education, and digital heritage presentation.
The development of image-based approaches for 3D dense reconstruction has found wide application in the architectural documentation of cultural objects and sites. For this reason, it is crucial to understand the potential and limitations of Multi-View Stereo (MVS) algorithms used in 3D shape reconstruction. This article aims to evaluate the quality and accuracy of dense point generation for cultural heritage objects using open-source (MicMac, OpenMVS, RealityScan), commercial algorithms (Agisoft Metashape and Pix4D), and an Unmanned Aerial Vehicle (UAV) in parallel and skew configurations at two historic wooden structures. To achieve this, a workflow for data evaluation was proposed, incorporating eigenvalues and quality factors, such as planarity, roughness, normal vector variance, and cloud-to-TLS (Terrestrial Laser Scanning) cloud deviation analysis. The results demonstrate that the image-acquisition geometry strongly influences reconstruction accuracy. Parallel configurations consistently produced higher completeness, narrower deviation distributions, and improved geometric fidelity compared to skew configurations. Among commercial solutions, Agisoft Metashape and Pix4D achieved the best overall performance. Open-source algorithms yielded variable results: OpenMVS-SGM achieved competitive accuracy under optimal conditions, while the OpenMVS-Patch-Based Approach offered a balance between density and roughness.
