Trainspotting

Trainspotting contains 12,068 objects and 311,886 photographs from Wikimedia Commons. Each object was photographed independently by many people over many years, in different environments and with different cameras. Each object is reconstructed with structure from motion and multi-view stereo, and every photograph comes with camera intrinsics and pose, multi-view stereo depth, and an object mask.

Paper: https://arxiv.org/abs/2611.06849

Project page: https://hanyuc.com/trainspotting/

Contents

Split	Objects	Photographs
train	11,068	279,319
eval	1,000	32,567

train/000/<id>.tar.gz ... train/002/<id>.tar.gz           one archive per object
eval/<id>.tar.gz                                          one archive per object
index.csv                                                 id, instance_id, split, path, num_images
regimes.csv                                               instance_id, split, regime
eval_samples.tsv                                          evaluation samples: instance_id, image_ids
covisibility.tar.gz                                       covisibility between the photographs of each object
sparse_final/train/000/<id>.tar.gz ... eval/<id>.tar.gz   sparse reconstruction of the released photographs
sparse_mvs/train/000/<id>.tar.gz ... eval/<id>.tar.gz     sparse reconstruction of the photographs used for multi-view stereo
license/license.parquet                                   per-photograph source and license
license/license_images_extra.parquet                      per-photograph source and license for images_extra
license/GFDL-1.2.txt                                      GNU Free Documentation License 1.2
viewer/v1/                                                assets of the interactive viewer

regimes.csv gives each object's regime: fixed is an individual installed at one site, roaming an individual that moves between sites, and instance-pooled a design shared by many physical instances.

<id> is the object's identifier: Q... is a Wikidata item, and C... is the page ID of a Wikimedia Commons category (https://commons.wikimedia.org/?curid=<number>).

Archive layout

Each archive extracts to one folder named by the object's instance_id:

<instance_id>/
  metadata.json
  images/<image_id>.jpg
  depths/<image_id>.npy
  masks/<image_id>.png

metadata.json:

instance_id     folder name, "<id>__<name>"
qid             object identifier (Wikidata item Q... or Commons category C...)
label           object name
num_images      number of photographs
images[]        one entry per photograph:
  id                       image_id
  image, depth, mask       paths relative to the object folder
  width, height            image size in pixels
  camera_model             PINHOLE (images are undistorted)
  intrinsics_px            fx, fy, cx, cy in pixels, for the image as stored
  camera_to_world_opencv   4x4 camera-to-world matrix, OpenCV convention (+x right, +y down, +z forward)

• Depth: float32 .npy, same size as the image, 0 where no depth is available.
• Mask: uint8 .png, same size as the image, 255 on the object.

eval_samples.tsv lists the 4- and 16-view evaluation samples on the eval split; image_ids are comma-separated, in evaluation order.

covisibility.tar.gz extracts to <instance_id>/covisibility.npz with image_ids and covisibility, a uint8 matrix whose entry [i, j] is the percentage of photograph i's masked depth points that project inside photograph j's object mask.

Sparse reconstructions

Each object has two COLMAP models (cameras.bin, images.bin, points3D.bin) with undistorted PINHOLE cameras, in the same frame and scale as metadata.json. sparse/final contains the released photographs. sparse/mvs contains the photographs used for multi-view stereo, including those in images_extra/, which are not in the main archives. All archives of an object extract into the same object folder:

<instance_id>/
  metadata.json, images/, depths/, masks/   main archive
  sparse/final/                             sparse_final archive
  sparse/mvs/, images_extra/                sparse_mvs archive

The image path of both models is the object folder.

License

The photographs come from Wikimedia Commons and keep their original licenses, which differ per photograph (mainly CC BY-SA and CC BY in various versions, public domain, and CC0; a small number under GFDL and other free licenses). license/license.parquet and license/license_images_extra.parquet list, for every photograph, the object (instance_id), the archive image (image_id), the Commons page ID and URL, the author, and the license. Any use of the photographs must attribute them according to their licenses. Photographs licensed under the GNU Free Documentation License are distributed under its terms; the license text is in license/GFDL-1.2.txt.

The annotations (camera parameters, depth maps, masks, sparse reconstructions, and metadata) are released under CC BY-SA 4.0.

Code and evaluation

The Spotter model, the evaluation protocol, and per-sample baseline results are at https://github.com/hanyucc/trainspotting. eval_samples.tsv defines the benchmark samples.

Citation

@misc{chen2026trainspotting,
  title = {{Trainspotting: Learning Object-Centric 3D Reconstruction from Found Photos}},
  author = {Chen, Hanyu and Snavely, Noah},
  year = {2026},
  eprint = {2611.06849},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2611.06849}
}