FlatLands — Generative Floormap Completion

One view is not one world

A single RGB frame reveals a narrow, occluded slice of traversable space. Several unseen layouts may explain exactly the same evidence, so a useful mapping system must preserve what it knows and represent what it cannot know from that view alone.

270,575observations

17,656real metric indoor scenes

6source datasets

From one view to a spatial belief

FlatLands separates perception from completion: first recover the visible floor in metric bird's-eye view, then model a distribution over the space beyond the camera frustum.

01 / Observe

Anchor the evidence in metric space

A virtual camera is placed at a physically valid pose in a reconstructed real scene. Visibility and ray occlusion determine exactly which floor cells the view supports in an aligned BEV grid.

Virtual camera observation synthesis from a reconstructed indoor scene to aligned BEV maps — Paper Figure 2b: offline observation synthesis used to construct FlatLands.

02 / Complete

Sample the unseen without rewriting the seen

At inference, one RGB frame yields depth and floor estimates, which are projected into observed BEV evidence. A conditional generator samples possible layouts while observed labels remain hard-clamped.

FlatLands pipeline from RGB input through depth and segmentation to BEV completion samples — Paper Figure 1: the complete RGB-to-floormap process.

Built from six real-world sources

The benchmark harmonizes distinct buildings, sensors, capture systems, and architectural styles into one metric BEV task. ScanNet++ is never seen during training and serves as the out-of-distribution test.

TUM + Google

3RScan

Multi-session RGB-D reconstructions of naturally changing indoor environments, with aligned textured meshes and consistent instance-level semantic annotations across rescans.

1,291FlatLands scenes

18,216retained observations

FlatLands input: OBJ meshes with semantic floor labels.

Project site Terms

Apple

ARKitScenes

Real-world mobile RGB-D captures collected with Apple LiDAR, including camera poses, surface reconstructions, registered depth, and labeled 3D furniture bounding boxes.

4,803FlatLands scenes

40,282retained observations

FlatLands input: PLY meshes with height-based floor extraction.

Project site Terms

Matterport + Academic Partners

Matterport3D

Building-scale RGB-D panoramas with globally aligned surface reconstructions, camera poses, and detailed 2D and 3D semantic segmentations across diverse real interiors.

2,101FlatLands scenes

38,004retained observations

FlatLands input: PLY meshes with semantic floor labels.

Project site Terms

TUM + Stanford + Princeton

ScanNet

A large RGB-D video corpus of indoor scans with recovered 3D camera poses, reconstructed surfaces, and instance-level semantic segmentations for scene understanding.

1,508FlatLands scenes

24,763retained observations

FlatLands input: PLY meshes with semantic floor labels.

Project site Terms

Zillow

ZInD

Panoramas from real unfurnished homes with annotated room layouts, doors, windows, openings, 3D camera poses, and final 2D floor plans produced through a human-in-the-loop pipeline.

7,026FlatLands scenes

133,096retained observations

FlatLands input: Panoramas with metrically calibrated floor geometry.

Project site Terms

TUM

ScanNet++

High-fidelity indoor reconstructions coupling sub-millimeter laser scans, registered DSLR imagery, iPhone RGB-D streams, and long-tail semantic annotations.

927FlatLands scenes

16,214retained observations

FlatLands input: PLY meshes; held out entirely for OOD testing.

Project site Terms

What one training example contains

Every observation keeps the input evidence, completion region, full target, valid workspace, and provenance aligned cell by cell. Evaluation is restricted to cells that are both valid and unobserved.

Ambiguity is the point

The four layouts below are real valid targets—not model samples. They agree on the shared observation and disagree only where the evidence permits different geometry. A calibrated model should cover those alternatives and localize uncertainty to their boundaries.

First valid ground-truth layout — Valid layout 1

Second valid ground-truth layout — Valid layout 2

Third valid ground-truth layout — Valid layout 3

Fourth valid ground-truth layout — Valid layout 4

Per-cell disagreement between valid layouts — Layout disagreement

What the benchmark reveals

Eleven methods—from training-free rules and deterministic networks to diffusion and flow matching—receive the same evidence and are scored under one uncertainty-aware protocol.

Best-of-4

Every stochastic generator surpasses every deterministic predictor when allowed to cover four plausible hypotheses.

Best MES

Flow Matching with Cross-Attention leads the multi-sample energy score on both in-distribution and held-out OOD scenes.

~600×

Its boundary variance is about 600 times its interior variance, concentrating uncertainty where room geometry is unresolved.

Use FlatLands

Download the validated release from Hugging Face and find evaluation code and baseline implementations on GitHub.

hf download Rudra1ssb/FlatLands FlatLands_final_dataset.zip --repo-type dataset
unzip FlatLands_final_dataset.zip -d FlatLands

Data use

FlatLands is a derived research dataset. The original captures and reconstructions remain governed by their source terms. Review LICENSES.md before use.

Cite FlatLands

If the dataset, benchmark, or models help your work, please cite the ECCV 2026 paper.

@inproceedings{bhattacharjee2026flatlands,
  title     = {{FlatLands}: Generative Floormap Completion From a Single Egocentric View},
  author    = {Bhattacharjee, Subhransu S. and Campbell, Dylan and Shome, Rahul},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2603.16016}
}

ECCV 2026 poster

Copyright & data rights

The derived BEV maps, masks, metadata, statistics, provenance records, and release text are provided under the FlatLands release notice. Underlying RGB-D captures, meshes, point clouds, panoramas, and source archives retain their original terms.

Paper figures are reproduced under the paper's CC BY 4.0 license. Source names, institution names, trademarks, and logos belong to their respective owners and identify provenance only.