16 June 2025 in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Recent advancements in aerial imaging, including high-resolution sensors and integrated GNSS/IMU systems, have significantly enhanced photogrammetric methods for geospatial data acquisition. While most aerial data is captured during daylight, night-time imaging is increasingly being used in applications such as urban analysis and disaster assessment. However, automatic co-registration of day and night imagery remains challenging due to substantial radiometric differences. This study investigates the use of deep learning-based feature matching techniques for the alignment of multi-temporal, day-night aerial datasets. Experimental results show that feature extraction is highly sensitive to scale, with only a limited subset of deep learning (DL) methods—particularly ALIKED with LightGlue and SuperPoint with SuperGlue—proving robust under low-illumination conditions. Additionally, a U-Net-like model was trained to pre-process night-time images by approximating their radiometric characteristics to those of daytime images, enabling consistent feature matching across all tested methods. Among them, ALIKED with LightGlue offered the best balance between match quantity and computational efficiency. Object-space evaluations confirmed that the proposed pre-processing step significantly improves co-registration accuracy. The methodology offers a promising foundation for future multi-sensor and multi-modal image alignment tasks, including RGB-thermal and 2D-3D matching.
28 July 2026 in International Journal of Computer Vision
16 June 2025 in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Image orientation, nowadays called Structure from Motion (SfM), is still an open research topic in particular in case of scenes featuring visual aliasing, or doppelgangers. Indeed, visually similar but distinct elements of the scene can cause incorrect matches, not detected by geometric or learning-based outliers removal methods, leading to misplaced camera poses and wrong 3D reconstructions. The paper reviews various state-of-the-art approaches to orient ambiguous image sequences and determination correct camera orientation parameters. We also present an in-house graph-based approach to reliably and precisely orient sets of images with doppelgangers. Different experiments on common ambiguous datasets are reported and commented.
3 November 2025 in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Road inspections are essential for assessing pavement conditions and planning maintenance to extend infrastructure lifespan. Cracks are key indicators of distress, and their timely detection can reduce repair costs. While manual inspections are costly and subjective, automated crack detection using vehicle-mounted cameras and digital image processing has improved efficiency but still requires extensive manual review. Crack detection research typically focuses on improving the accuracy of the detection and feature extraction, neglecting the geospatial mapping part which is essential in pavement inspection. This paper presents an end-to-end pipeline that combines V-SLAM, deep learning (DL) and 2D-3D mapping to detect and map pavement cracks using vehicle-mounted cameras footage and provide georeferenced results interpretable inside a Geospatial Information System (GIS). Results significantly reduce inspection time by providing preliminary detections for rapid verification by expert operators in GIS environments. This enhances data management, efficient validation, advanced spatial analysis and time-based tracking of crack progression, ensuring informed decision-making and optimized maintenance planning, ultimately extending infrastructure lifespan and reducing costs.