SceneAligner data

Data of SceneAligner: 3D-Grounded Floorplan Localization in the Wild (code, model): our clean subset of C3 and cached data of its test split. The photos, floorplans and annotations come from C3.

Path	Content
train.csv, test.csv	Our subset of C3: one row per photo-floorplan pair (columns described in the code repository)
test/reconstructions/<scene_id>_<chunk_idx>.npz	π³ reconstruction of each test chunk (at most 150 photos), as used in the paper
test/density_maps/<scene_id>_<chunk_idx>.png	Density map of each reconstruction, as computed by the evaluation (512 × 512; confidence 45 %, x/z 2.5 %, heights 20-95 %, gamma 0.5)
test/geocalib.csv	GeoCalib gravity of each test photo

Reconstructions (NumPy .npz):

Key	Shape, type	Content
uid	(S,) int64	Photos, in the order of test.csv
points_cam0	(S, 518, 518, 3) float32	3D point of each pixel of the photos resized and padded to 518 × 518, in the frame of the first camera (OpenCV convention: x right, y down, z forward)
conf	(S, 518, 518) float32	Confidence of each point
extrinsics	(S, 4, 4) float32	Camera-to-first-camera poses
est_gravity_cam	(3,) float32	Gravity direction of the scene in the frame of the first camera: spherical medoid of the GeoCalib gravity of the photos

test/geocalib.csv: uid, gravity_x, gravity_y, gravity_z, the gravity direction (pointing down) in the camera frame of the photo after its EXIF orientation (OpenCV convention).

With the code repository, python evaluate_c3.py --c3_root <C3> --recon_cache <this dataset>/test/reconstructions reproduces Tables 1 and 2 of the paper.

License

Derived from C3 and released under the same license, CC BY 4.0.

Citation

@inproceedings{cho2026scenealigner,
  title={SceneAligner: 3D-Grounded Floorplan Localization in the Wild},
  author={Junhyeong Cho and Ruojin Cai and Hadar Averbuch-Elor},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}