Neural understanding of objects and materials

This Thesis addresses fundamental challenges in neural understanding of objects and materials through three research domains: depth estimation for non-Lambertian surfaces, anomaly detection and segmentation in advanced industrial settings, and modern image forgery detection and localisation in the era of sophisticated generative models. The work demonstrates that effective visual understanding in complex real-world scenarios requires specialised methodological approaches that explicitly account for domain-specific challenges rather than relying on general-purpose solutions. The first part of this research tackles the long-standing challenge of accurate depth estimation for non-Lambertian surfaces that violate fundamental assumptions underlying traditional computer vision methods. A novel benchmarking framework and a learning-based approach are developed that enable robust geometric perception for optically complex materials through innovative training strategies and comprehensive evaluation methodologies. In the second part, the Thesis introduces methodological innovations that address fundamental limitations in existing unsupervised approaches for industrial anomaly detection and segmentation. Crossmodal feature mapping paradigms, test-time training strategies, and lightweight architectures are developed to enable practical deployment while achieving superior performance across diverse industrial inspection scenarios. Finally, the introduction of a comprehensive multiview multimodal benchmark reveals significant gaps in current algorithms while establishing an evaluation framework for future progress. Finally, in the third part, contemporary challenges posed by sophisticated diffusion-based manipulation techniques are addressed in the context of image forgery detection and localisation. Foundation model features are combined with multimodal information fusion strategies to achieve reliable identification of high-quality image manipulations, demonstrating how principles developed in anomaly detection can be adapted to distinguish authentic from manipulated digital content. The thesis unifies three research domains through shared methodological principles. Specialised learning consistently outperforms general approaches, with multimodal fusion and computational efficiency guiding robust, practical designs. Contributions include novel algorithms, datasets, training strategies, and evaluation frameworks, validated through extensive experiments establishing effectiveness and exposing future research.