Reconstructing Interiors From a Handful of Photos

Most of the buildings we will live and work in over the coming decades already exist. Renovating them, selling them, furnishing them or simply managing them all start from the same question: what does this space actually look like, and what are its dimensions? For a large share of homes, the honest answer is a few photos and, at best, an outdated plan. A faithful 3D model of a room, a digital twin, turns that question into data. In the long run, it enables: - Real estate: remote visits and floor plans straight from the listing photos. - Renovation and energy retrofit: as-built geometry to plan works and quantities. - Interior design and home staging: trying real furniture, at real dimensions, before buying. - Architecture and BIM: a starting model instead of a manual survey. - Insurance and facility management: a documented state of a space at a given date. - Spatial computing: a measured model of the room for augmented reality and robotics. None of these uses needs a scan of a museum piece. They need many ordinary rooms, digitised cheaply, at scale.

Three families of methods exist today. Each fails one of the requirements above: low cost, few photos, faithful measurements, or an editable result. 2.1 Dedicated sensors are accurate but expensive. Laser scanners and LiDAR give excellent geometry. They also require hardware, a visit on site and an operator. That is acceptable for a landmark building, not for millions of apartments. 2.2 Photogrammetry needs far more photos than a listing provides. Classic photogrammetry reconstructs geometry from photos alone, but its cost and data needs grow quickly with the number of images.

Photogrammetry also needs strong overlap between views and textured surfaces to match. Its known failure cases are plain white walls, glass and reflections, and sparse photo sets: exactly what a listing interior looks like. 2.3 Neural and generative methods look right but are not measured. Neural rendering (NeRF, Gaussian Splatting) fits millions of parameters per scene. In our trials, one room took 30 to 69 minutes, looked stunning from the input photos, degraded away from them, and contained no separable walls or furniture. General-purpose AI agents and generative world models produce plausible rooms. In our 3-photo test, an agent read the room well, but its walls were off by 0.5 to 1 m. 2.4 Two fundamental difficulties. - Scale. Photos alone do not contain absolute scale: the same images are consistent with a room and with a model of it at any size. Without a sensor, real dimensions must be inferred. - Unseen areas. With a few photos, parts of the room are seen once or never: behind furniture, or on the photographer's side. They must be completed without inventing what is not there.

Research question: can we obtain a measured, faithful and editable model of a room from the 3 to 9 photos an ordinary listing already provides, on affordable hardware?

We build a structured reconstruction: the output is not a cloud of points or a picture, but a space made of walls, openings and objects, with real dimensions. That structure is what makes the result usable: it can be measured, exported to BIM (IFC), furnished, edited and explored in a web browser.

We evaluate on real rooms along four axes, keeping internal consistency separate from accuracy against the real world: geometric consistency, absolute accuracy against a tape measure, generalisation to unseen viewpoints, and compute time. 4.1 Geometric consistency: 5 rooms, 3 to 9 photos

This deviation measures how consistent the walls are with the geometry reconstructed from the photos. It is not an absolute, centimetre-level accuracy. The structure is stable wherever the photos see the walls, including walls that are not at right angles. Coverage matters more than photo count: 3 photos are enough for a studio, not for a 37 m² living room. 4.2 Accuracy against reality: a first ground truth We measured a few dimensions with a tape measure in a 20 m² room reconstructed from 9 photos.

With no on-site measurement at all, the two structural dimensions deviate by 3.4% and 3.8%, which bears directly on the scale problem of section 2.4. The window and the desk are clearly less accurate. A single reference dimension, such as the ceiling height, could improve the overall scale estimate. Only one room has been surveyed so far: a first data point, not yet a statistic. 4.3 Generalisation to unseen viewpoints Judging a scene from the photos that built it is the easiest case. For each photo, we rebuild the room without it, then compare the scene with that photo. Across two rooms (15 tests), the loss on an unseen viewpoint stays small: 0.2 to 1.4 dB of PSNR. The same test shows that removing a photo can still change the structure: capture quality remains decisive. 4.4 Compute time

The results support the research question for single rooms with good photo coverage. They also mark clear limits: - Ground truth is still thin. One room surveyed with a tape measure; more are needed to quantify accuracy statistically. - Coverage drives robustness. Removing one photo still changes the structure in some tests; a simple capture protocol is in progress. - One room at a time. Whole homes, with navigation between rooms, remain to be done. - Furniture is approximate. Right place and colour, but catalogue shapes rather than the exact pieces photographed.

Once a room is measured and structured, it becomes a workspace. Furniture from a real catalogue, with real dimensions and prices, can be placed, swapped or restyled, and an assistant can take plain-language requests while respecting the room's constraints. This is 3D home staging based on real assets, rather than on generic look-alikes.

A new phase is opening for our R&D. With access to high-performance computing (HPC) resources, we will be able to speed up our experiments, explore new optimisations and work on reconstructing entire homes. All the figures above were obtained on a single consumer graphics card. The gains from HPC will depend on parallelisation and software optimisation: we will measure them before announcing them.

Our long-term ambition: drastically reduce the time it takes to turn a few photographs into interactive 3D spaces, and move towards a near real-time experience.

A handful of listing photos already contains most of what is needed to rebuild a room. Our first results show that a structured approach can recover centimetre-consistent walls and dimensions within a few percent, in the measured ranges above, on affordable hardware. The next steps are more ground truth, whole homes, and much shorter compute times.

This article reports results from our own trials; it does not describe our method. The background it draws on is covered in these works: - Schönberger, J. L. and Frahm, J.-M. (2016). Structure-from-Motion Revisited. CVPR 2016. (COLMAP) - Triggs, B., McLauchlan, P., Hartley, R. and Fitzgibbon, A. (1999). Bundle Adjustment — A Modern Synthesis. - Mildenhall, B. et al. (2020). NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV 2020. - Kerbl, B., Kopanas, G., Leimkühler, T. and Drettakis, G. (2023). 3D Gaussian Splatting for Real-Time Radiance Field Rendering. SIGGRAPH 2023.