Mirrors as Free Extra Cameras for Sparse-View, Metric-Scale Reconstruction

When a camera sees a flat mirror, it is really seeing the scene twice: once directly, and once from a "virtual camera" standing behind the glass. Today's 3D reconstruction pipelines treat that second view as a problem. Monocular depth networks, which predict depth from one image, interpret the reflection as a room behind the wall. Radiance-field methods place fake geometry there. Methods that handle mirrors correctly need hundreds of photos. We propose the opposite stance for the regime where photos are scarce, meaning two to five views. There, the depth a monocular network predicts inside a mirror is taken as a correct measurement of virtual geometry. It is reflected across an estimated mirror plane and used to supervise the real scene from a viewpoint no physical camera occupied. Each real photo containing a mirror then contributes a second, perfectly synchronized view with identical optics. Sparse-view reconstruction is short of exactly this kind of multi-view evidence.

The same reflection geometry gives a second, smaller benefit. Metric depth networks only produce meaningful meters when told the camera's focal length. We show how an oblique mirror lets us recover focal length from the image alone. We are careful about what mirrors cannot do. A mirror does not fix absolute scale, because scaling the scene, the mirror and its distance together leaves every image unchanged. Absolute scale must still come from the learned prior of the metric depth network. What the mirror supplies is the camera calibration that prior needs, and a second, independent observation of every reflected object from which to estimate it.

Core Claim

In sparse-view reconstruction of scenes containing planar mirrors, three signals, combined, should beat both mirror-agnostic sparse-view methods and dense-view mirror methods run at low view counts. The first is a mirror plane estimated per image from metric monocular depth, rather than from the splat cloud. The second is reflected monocular depth used as virtual-view supervision. The third is focal length self-calibrated from real/reflected correspondences. The gain should be largest on geometry visible only through the mirror.

The mechanism rests on one relation. A real 3D point $X$ and its mirror image $X'$ satisfy

$$X' = (I - 2nn^\top)X + 2dn,$$

where $n$ is the mirror's unit normal and $d$ its distance from the camera. Each point is back-projected as $X = D(u)K^{-1}[u;1]^\top$, where $D(u)$ is the predicted metric depth at pixel $u$ and $K$ holds the focal length $f$. Given matched real and reflected pixels, we solve for $f$, $n$ and $d$ by least squares on this relation.

A wrong $f$ stretches the back-projected scene laterally relative to depth. That stretch commutes with reflection when $n$ is parallel or perpendicular to the optical axis, so $f$ is only identifiable for oblique mirrors. This is a sharp prediction we test directly. A global scale on $D$ rescales $d$ with it and is never identifiable from the relation.

Where This Sits

Three lines of prior work motivate the proposal by exposing the gap.

Reflect3r flips the detected mirror region of a single image to fake a second view and feeds the pair to the DUSt3R stereo network. Using a symmetry loss based on the reflection transform, it reaches 89.37% completeness against 79.50% for the monocular MoGe baseline on a 16-scene synthetic benchmark. This is the strongest evidence that reflections carry usable stereo geometry. However, the method works in an unscaled coordinate frame, stops at a point map, and never builds a renderable scene from several views.

Mirror-3DGS and Gaussian Splatting in Mirrors render each frame twice, from the real camera and its mirrored twin, inside Gaussian splatting (a scene made of many small colored 3D blobs rendered by projection). Both find the mirror plane by random-sample plane fitting (RANSAC) on an already-trained dense reconstruction, using 200–300 views per scene. At three views that reconstruction is full of floaters, which are spurious semi-transparent blobs, so the plane fit has nothing stable to fit.

Mirror3D goes the other way. It estimates the mirror plane from depth in a band around the mirror frame and overwrites the reflection with a flat surface, cutting mirror depth RMSE on NYUv2 from 1.170 to 0.874. It throws away the reflected content.

Our proposal borrows its border-band plane estimate as an initialization but keeps and reuses exactly what Mirror3D discards.

Two other groups of work constrain the design. On the sparse-view side, Few-Shot Depth-Aware 3DGS supervises splats with monocular depth under a scale-invariant loss. It also lowers the color model's complexity and disables pruning to avoid memorization. We adopt these tricks as our base, but that paper trusts in-mirror depth as real, which would cement a ghost room. RegNeRF and StructNeRF regularize unseen views and planar regions. StructNeRF explicitly reports that its photometric consistency terms fail on specular and mirror surfaces. MVSplat predicts splats feed-forward from a plane-sweep cost volume, a stack of cross-view matching scores at candidate depths. That cost volume assumes surface color does not change with viewpoint, which mirrors violate.

On the scale side, Metric3D v2 maps images to a canonical focal length to predict meters (NYUv2 AbsRel 0.047). It depends on correct focal metadata and misreads mirrors as open space. Our self-calibration targets precisely that dependency.

Research Program

The system has four stages.

Stage 1: per-image mirror geometry. For each sparse input image we obtain a mirror mask with the segmenter used by Reflect3r or with Mirror3DNet. We predict metric depth with Metric3D v2, and with UniDepth or Depth Pro as alternatives. We initialize $n$ and $d$ from the border band, following Mirror3D. We then match real content to flipped in-mirror content, as Reflect3r does, and solve the reflection relation for $n$, $d$ and, when metadata is absent, $f$. The output is a per-view mirror plane together with a reflected depth map: in-mirror depths mapped into real-world points seen from the virtual camera.

Stage 2: fuse the plane across views. The per-view planes are expressed in the shared world frame using the given camera poses and fused robustly. Even three estimates from independent images are far more reliable than a plane fitted to a sparse, floater-ridden splat cloud. A single global scale factor is fitted jointly over all real and reflected depth observations.

Stage 3: sparse Gaussian splatting with virtual cameras. Each real camera spawns its mirrored virtual camera, and the two renders are composited by the mirror mask, following Mirror-3DGS. Geometry is supervised by metric depth outside mirrors and by reflected depth inside them. The mirror plane is refined photometrically, following Gaussian Splatting in Mirrors, but starting from the Stage-2 estimate rather than from RANSAC.

Stage 4: a deliberately modest cross-modal component. Instead of a hard 2D mask, each Gaussian carries a mirror probability. It is obtained by projecting the Gaussian into every input view, with a visibility check, and averaging the segmenter's per-pixel mirror scores. This lets view-inconsistent mask errors cancel. We treat it as a hypothesis to test, not a central contribution. SGFormer-style cross-attention from 3D points to 2D semantic tokens is the heavier variant we will try only if simple projection-averaging leaves measurable mask-boundary errors.

Baselines

Naive reference. The do-nothing point is vanilla Gaussian splatting on the same two, three and five views. It shows how badly mirrors and sparsity combine.

Strongest mirror-agnostic sparse method. This is Few-Shot Depth-Aware 3DGS, run with the same metric depth network, the same pose inputs and the same iteration budget. It is the closest method that already exploits monocular depth under sparsity. On LLFF it reports relative gains of +10.5% PSNR and +14.1% LPIPS over its baseline. The file does not report its absolute numbers on mirror scenes, and we found no reported result on Lacuna for this method on the Mirror-NeRF data. Beating it means higher PSNR inside mirror regions and lower depth error on real surfaces, without losing PSNR outside mirrors.

Strongest mirror-aware methods. These are Mirror-3DGS and Gaussian Splatting in Mirrors, run on the identical subsampled views with their released schedules, with iteration counts matched to ours. At full density, Gaussian Splatting in Mirrors reports 39.87 PSNR on Mirror-NeRF synthetic scenes against 36.52 for vanilla splatting. We treat 39.87 as the dense-view ceiling. We found no reported result on Lacuna for either method at five or fewer views. Beating them means a statistically clear PSNR and mirror-plane-angle advantage at three views, averaged across scenes and view samplings.

Reflection-stereo initialization. This baseline runs Reflect3r on each image and initializes splats from the merged point maps. The merge uses a per-view scale alignment, since Reflect3r's outputs are unscaled. It is the most direct competitor to our Stage 1.

Feed-forward reference. MVSplat is run zero-shot from its RealEstate10K checkpoint. For context, it reports 26.39 PSNR on RealEstate10K.

Ablations. Three ablations isolate the key ingredient. The first removes reflected-depth supervision while keeping virtual rendering. The second replaces the Stage-2 plane with RANSAC on splats. The third compares EXIF focal, self-calibrated focal, and focal deliberately perturbed by ±20%.

Why not just…?

Why not just mask mirrors out, fill them with Mirror3D's plane, and run a good sparse-view method? This is the most serious threat. For content also seen directly, reflections may add little. The decisive evaluation therefore splits scenes into regions visible directly and regions visible only through the mirror. If our gains appear only in mirror pixels and vanish for directly visible geometry, the honest conclusion is that masking suffices for most uses. We would then report the contribution as "recovering mirror-only geometry." Mirror3D's own design discards that geometry by construction, so it cannot recover it.

Why not just run Mirror-3DGS or Gaussian Splatting in Mirrors on few views? Their plane estimate is fitted to the splat cloud. Under sparse views that cloud is dominated by floaters, as RegNeRF and Few-Shot Depth-Aware 3DGS both document. This is still a prediction rather than a measured result. The plane-source ablation tests it directly: if their pipeline with our Stage-2 plane swapped in matches our full system, then plane initialization is the whole story and reflected-depth supervision is not.

Why not just use EXIF focal length and skip self-calibration? When EXIF is present and correct, one should. Self-calibration only earns its place for stripped or corrupted metadata, which Metric3D v2 is known to depend on. We will report metric depth error on the Mirror3D frames under three conditions: true focal, perturbed focal, and self-calibrated focal. We will also report how calibration error varies with mirror obliqueness.

Why not just use a 2D mask instead of per-Gaussian semantic lifting? The mask may be enough. The ablation compares hard single-view masks with multi-view-averaged mirror probabilities. It uses boundary depth error around mirror frames as the metric, since this is where Reflect3r notes mask noise corrupts geometry. If there is no difference, the component is dropped.

Proposed Experiments

The main benchmark is the Mirror-NeRF dataset: three synthetic and three real scenes, subsampled to two, three and five views, with five random view draws per setting. We report PSNR, SSIM and LPIPS on held-out views, computed separately inside and outside mirror masks, plus mirror-plane normal angular error. On the synthetic scenes, which have depth, we also report depth error. For reference, Mirror-3DGS reports +3.38 dB over vanilla splatting on the real scenes at dense views.

Geometry is also evaluated on the Reflect3r synthetic benchmark, using completeness and accuracy against ground-truth virtual space, where Reflect3r reports 89.37% and 96.64%. For metric scale we use ScanNet and Matterport3D sequences that contain the Mirror3D-annotated mirror frames, taking three to five frames each with dataset poses. Refined mirror depth provides ground truth, and we report AbsRel and RMSE in meters. For context, Mirror3DNet reports mirror RMSE of 0.874 on NYUv2. To check that the method does no harm on ordinary scenes, we run LLFF with five views.

Decisive check 1: mirror-only geometry. Depth accuracy on geometry visible only through the mirror must improve over the mask-and-fill baseline. If it does not, the core claim fails.

Decisive check 2: the obliqueness curve. Self-calibrated focal error should be smallest for mirrors near 45° to the optical axis and diverge as the normal approaches either the parallel or the perpendicular configuration. We test this by sweeping synthetic mirror orientations. Failure to see this curve would mean the solver is fitting something other than the reflection geometry.

Decisive check 3: scale behavior. Absolute metric error should track the depth network's own scale error and not be removed by the mirror. We expect only a variance reduction from averaging real and reflected estimates. This check keeps the scale claim honest.