Reflect3r: Single-View 3D Stereo Reconstruction Aided by Mirror Reflections

TL;DR

Reflect3r addresses the challenge of single-view 3D reconstruction in scenes with mirrors by reframing mirror reflections as auxiliary views and constructing a physically valid virtual camera in pixel space. The approach integrates with existing feed-forward multi-view models (e.g., DUSt3R) and introduces a symmetric-aware loss to enforce geometric consistency between real and virtual poses, with an extension to dynamic scenes. A fully synthetic, editable Blender dataset with ground-truth real and virtual poses supports quantitative evaluation. Experiments on real and synthetic data show that Reflect3r achieves higher scene completeness, better accuracy, and lower Chamfer distances than strong baselines, demonstrating the practical value of leveraging mirror-induced stereo cues. The work provides a reusable framework and dataset for robust, low-cost 3D reconstruction in unconstrained environments.

Abstract

Mirror reflections are common in everyday environments and can provide stereo information within a single capture, as the real and reflected virtual views are visible simultaneously. We exploit this property by treating the reflection as an auxiliary view and designing a transformation that constructs a physically valid virtual camera, allowing direct pixel-domain generation of the virtual view while adhering to the real-world imaging process. This enables a multi-view stereo setup from a single image, simplifying the imaging process, making it compatible with powerful feed-forward reconstruction models for generalizable and robust 3D reconstruction. To further exploit the geometric symmetry introduced by mirrors, we propose a symmetric-aware loss to refine pose estimation. Our framework also naturally extends to dynamic scenes, where each frame contains a mirror reflection, enabling efficient per-frame geometry recovery. For quantitative evaluation, we provide a fully customizable synthetic dataset of 16 Blender scenes, each with ground-truth point clouds and camera poses. Extensive experiments on real-world data and synthetic data are conducted to illustrate the effectiveness of our method.

Figures (11)

• Figure 1: Single-view 3D reconstruction with mirror reflections. Given an image containing a mirror, we aim to reconstruct the 3D geometry of the scene. Existing methods cannot recognise the reflective cues and fail by predicting a false geometry for the mirror region, which is highlighted with light red (Top). We reinterpret the mirror reflection as a virtual view captured by a simulated camera, enabling a stereo formulation that leads to more accurate geometry reconstruction (Bottom).
• Figure 2: Physical imaging process of a scene containing a mirror. The reflection plane is shown as semi-transparent to reveal the virtual camera.
• Figure 3: Overview of the proposed Reflect3r pipeline. Reflect3r reconstructs 3D scenes from a single-view image by leveraging mirror reflections. A reflection transformation is designed to ensure that flipping the real view in the pixel domain, simulating a virtual camera imaging, enables seamless integration with modern feed-forward models. Following the initial prediction, the reflection transformation is used as a geometric constraint to refine pose optimization.
• Figure 4: Thumbnails of the dataset, where each image represents a fully customizable Blender scene.
• Figure 5: Qualitative results of Reflect3r and all the baselines running on the real-world data, where the predicted geometries corresponding to the mirror area are highlighted with light red.