Advancing mirror segmentation with spatio-temporal and depth cues

Xu, Mingchen 2025. Advancing mirror segmentation with spatio-temporal and depth cues. PhD Thesis, Cardiff University. Item availability restricted.

	PDF (Mingchen Xu, PhD Thesis) - Accepted Post-Print Version Restricted to Repository staff only until 19 February 2027 due to copyright restrictions. Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (11MB)
	PDF (Cardiff University Electronic Publication Form) Restricted to Repository staff only Download (216kB)

Abstract

Mirror segmentation aims to identify and localize mirror regions in complex scenes, a task challenged by reflective ambiguity, unclear boundaries, and misleading contextual cues. While related to glass, shadow, and salient object segmentation, mirrors pose unique difficulties due to virtual content and inconsistent visual or depth cues. This thesis presents a comprehensive study of mirror segmentation across image, video, and RGB-D modalities. We first review related segmentation tasks to analyze how their principles inform mirror detection. Building on this foundation, we propose a video mirror segmentation framework that jointly models short term motion inconsistencies and long-term contextual relations. Observing similar local-global patterns in depth, we design a depth-guided mirror detection module based on a triple consistency principle, enforcing alignment between spatial, semantic, and geometric cues. We also introduce a large-scale RGB-D video mirror dataset covering diverse real-world scenes. Experiments show that our method surpasses previous state-of-the-art approaches on multiple benchmarks. Motivated by the success of vision foundation models, we further adapt SAM2 for mirror understanding, developing MirrorSAM2, which integrates SAM2’s generalization ability with frequency-domain priors. This enables more accurate distinction of mirror regions from complex reflections in both image and video settings. Overall, this thesis advances reflection-aware scene understanding and provides a unified framework for handling reflective surfaces. Future research may further integrate frequency priors, generative diffusion models, and foundation architectures for more robust mirror perception

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Date of First Compliant Deposit:	19 February 2026
Date of Acceptance:	18 February 2026
Last Modified:	19 Feb 2026 11:10
URI:	https://orca.cardiff.ac.uk/id/eprint/185015

Actions (repository staff only)

Edit Item