FlatLands

Generative Floormap Completion From a Single Egocentric View

Subhransu S. Bhattacharjee · Dylan Campbell · Rahul Shome

TL;DR: FlatLands asks a model to complete a metric indoor floormap from one partial egocentric observation and represent multiple plausible hidden layouts instead of forcing a single guess.

_{A physically valid virtual camera observation is back-projected, rasterized, and aligned into egocentric BEV maps.}

_{One RGB frame becomes observed BEV evidence, then a conditional generator samples plausible hidden layouts.}

• Dataset archive and Hub preview are live on Hugging Face.
• Model weights, construction code, and additional benchmark tooling are planned for release in this repository.
• This repository does not redistribute upstream RGB-D captures, meshes, point clouds, panoramas, or source archives.

Source	What it contributes to FlatLands
3RScan	Changing indoor environments with aligned multi-session RGB-D reconstructions and semantic OBJ meshes. 1,291 scenes / 18,216 observations.
ARKitScenes	Mobile LiDAR RGB-D captures with poses, reconstructed PLY surfaces, registered depth, and labeled furniture. 4,803 scenes / 40,282 observations.
Matterport3D	Building-scale RGB-D panoramas, globally aligned reconstructions, camera poses, and semantic PLY meshes. 2,101 scenes / 38,004 observations.
ScanNet TUM + Stanford + Princeton	Indoor RGB-D scans with recovered camera poses, surface reconstructions, and instance-level semantic PLY meshes. 1,508 scenes / 24,763 observations.
ZInD	Panoramas of real homes with room layouts, openings, camera poses, floor plans, and metric floor geometry. 7,026 scenes / 133,096 observations.
ScanNet++ TUM	High-fidelity laser scans, DSLR imagery, iPhone RGB-D, and long-tail semantics; reserved for OOD testing. 927 scenes / 16,214 observations.

Counts are FlatLands scene layouts and retained observations, not the upstream datasets' published totals. See LICENSES.md for source terms.

Each observation contains four aligned binary maps plus metadata. The evaluation region is the valid unobserved area, so observed evidence is preserved and only the hidden floor map is completed.

RGB context

Observed floor

Unobserved mask

Complete floor map

Validity mask

obs_*/
  observed_floor.png
  floor_map.png
  unobserved.png
  epistemic_mask.png
  metadata.json

FlatLands treats floormap completion as a posterior prediction problem. A single partial observation may be consistent with several valid room layouts; benchmark metrics therefore include both fidelity and multi-sample uncertainty.

Observation

Completion 1

Completion 2

Completion 3

Completion 4

Variance

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

Archive integrity:

The Hub dataset viewer contains small preview parquet splits; the full 270,575-observation release is in the archive above.

FlatLands is a derived research dataset. Users must comply with the upstream source dataset terms listed in LICENSES.md. If an upstream term is more restrictive than this release notice, the upstream term controls for the observations derived from that source.

Copyright (c) 2026 Subhransu S. Bhattacharjee, Dylan Campbell, and Rahul Shome. FlatLands release materials, derived BEV maps, masks, metadata, statistics, and provenance records are provided under the FlatLands release notice in LICENSE. The website and README media are research figures for explaining the benchmark. Paper figures are copyright (c) 2026 the FlatLands authors and reproduced under CC BY 4.0; underlying source dataset assets remain governed by their original terms. See COPYRIGHT.md and LICENSES.md.

@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}
}

Please also cite the relevant upstream datasets for any FlatLands observations used in your work.