Ahmad, Waqar (2026) Weakly Supervised Roadside Object Geolocation from Street-Level Imagery with Similarity Learning and Graph-Based Reasoning. PhD thesis, Dublin City University.
Abstract
Mapping roadside infrastructure from street-level imagery is important for urban asset management and smart-city applications. Large collections of street-level images enable automated infrastructure monitoring. However, estimating the real-world locations of roadside objects remains challenging due to weak supervision, noisy detections, and uncertain geometric measurements across viewpoints. In practice, infrastructure records are often incomplete or outdated. Obtaining detailed annotations for large image datasets is costly and difficult to scale. This thesis investigates methods for detecting and geolocating roadside objects from street-level imagery under weak supervision. The aim is to enable reliable object localization under noisy and uncertain conditions. The proposed framework integrates visual localization, similarity learning, and geometric reasoning to address these challenges. Roadside objects are first localized in images using weak supervision. Multi-view geometric cues are then used to estimate their real-world locations from multiple camera observations. However, weak supervision and triangulation can introduce noisy observations and incorrect associations across views. Similarity learning is applied to estimate visual similarity between image pairs. A statistical outlier detection framework identifies and filters unreliable training pairs. This improves the reliability of the resulting geolocation estimates. To move beyond purely statistical inference, a learning-based graph model is introduced. This model improves the robustness and scalability of the geolocation process. The model represents relationships between observations and candidate object locations using a graph neural network. It learns to identify consistent multi-view evidence while suppressing spurious geometric intersections from noisy measurements. Extensive experimental evaluation demonstrates that the proposed methods achieve competitive and robust roadside object geolocation, even when observations are noisy and geometric measurements are uncertain. Overall, this thesis develops and validates methods for detecting and geolocating roadside infrastructure from street-level imagery. This contributes to more accurate and reliable automated mapping of urban assets.
Metadata
| Item Type: | Thesis (PhD) |
|---|---|
| Date of Award: | August 2026 |
| Refereed: | No |
| Supervisor(s): | Krylov, Vladimir |
| Subjects: | Mathematics > General Mathematics Mathematics > Mathematics (miscellaneous) Mathematics > Analysis |
| DCU Faculties and Centres: | DCU Faculties and Schools > Faculty of Science and Health > School of Mathematical Sciences |
| School, Faculty or Unit: | Faculty of Science & Health Faculty of Science & Health > School of Mathematical Sciences |
| Use License: | This item is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 License. View License |
| Funders: | ADAPT Centre |
| ID Code: | 33278 |
| Deposited On: | 09 Nov 2026 15:45 by Vladimir Krylov . Last Modified 09 Nov 2026 15:45 |
Documents
Full text available as:
Preview |
PDF
- Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Creative Commons: Attribution-Noncommercial-No Derivative Works 4.0 14MB |
Downloads
Downloads
Downloads per month over past year
Archive Staff Only: edit this record