Abstract
dc:descriptionMonocular Depth Estimation (MDE) aims to infer a dense depth map from a single image. Because a three-dimensional scene is reduced to a two-dimensional projection, the inverse mapping is fundamentally ill-posed, since multiple geometries can be consistent with the same image evidence. As a result, high benchmark performance may coexist with solutions that are difficult to audit with respect to the depth cues on which the model relies, the internal organization of its computations, and the regions where failures concentrate. This dissertation approaches MDE through three complementary pillars: interpretability, performance, and explainability. Under the first pillar, it introduces the SVDA attention mechanism, inspired by singular value decomposition, which factorizes attention into directional alignment and spectral emphasis, enabling the study of its structure through quantitative indicators. Under the second pillar, it investigates a segmentation-based approach in which compact class-conditioned depth predictors are combined to improve depth discontinuities at semantic boundaries. Under the third pillar, it proposes a cue-based explainability framework and the Visual Depth Cue (VDC) dataset for the controlled evaluation of model behavior with respect to human visual depth cues. The results show that SVDA preserves competitive predictive performance while revealing structured and analyzable attention behavior, that semantically aware decomposition improves prediction quality at object boundaries, and that controlled experiments on VDC reveal a gap between benchmark performance and cue-consistent behavior. The study focuses on static monocular images and Transformer-based architectures, leaving the investigation of temporal sequences, stereo constraints, and robustness under imperfect segmentation as directions for future work.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Arampatzakis, Vasileios
- Αραμπατζάκης, Βασίλειος
Subjects
dc:subject × 16- Μονοσκοπική Εκτίμηση Βάθους
- Ερμηνευσιμότητα
- Επεξηγησιμότητα
- Μετασχηματιστές Όρασης
- Μηχανισμοί Προσοχής
- Monocular Depth Estimation
- Interpretability
- Explainability
- Vision Transformers
- Attention Mechanisms
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Τεχνολογία μέσων
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Media Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61501