MVMD detects mirrors across views, boosting 3D reconstruction accuracy by 11.1% IoU

Mirrors wreak havoc on 3D reconstruction: they create distorted, fragmented spaces that make photogrammetry and neural radiance fields produce inaccurate geometry. Traditional mirror detection only looks at single images, missing the rich correspondence clues available when the same scene is captured from multiple angles. To solve this, Yidan Shen and colleagues from the computer vision community propose MVMD (Multi-View Mirror Detection), a novel architecture that explicitly models how objects shift inside mirrors as the camera moves, distinguishing reflections from real geometry.

MVMD is built from three specialized blocks: an Inter-Views Block that tracks viewpoint-dependent movement of mirrored objects, an Intra-View Block that detects reflections within a single frame, and a Refinement Block that sharpens mirror boundaries. These blocks are trained jointly using cross- and self-attention mechanisms, letting the model associate features across views and across the mirror edge. The authors also contribute the first dataset purpose-built for multi-view mirror detection. In experiments, MVMD outperforms single-image detectors by up to 2.6% in accuracy and 11.1% in Intersection-over-Union (IoU), a substantial gain for downstream tasks. The work, published at WACV 2025, is now available on arXiv, making it more accessible to the research community.