Trainspotting

Learning Object-Centric 3D Reconstruction from Found Photos

Hanyu Chen Noah Snavely

Cornell University

12,068 objects · 311,886 photographs

Project page Film credits
Trainspotting

Trainspotting: Learning Object-Centric 3D Reconstruction from Found Photos

Hanyu Chen Noah Snavely

Cornell University

Explore all 12,068 objects

Abstract

State-of-the-art feedforward reconstruction models perform poorly on in-the-wild object-centric photo collections. They are trained largely on single-session object captures whose images vary in viewpoint while the object and its environment remain nearly fixed. Public photo archives contain a different form of multi-view data. They hold found photos of the same object or design: photos taken independently by many people across years, environments, cameras, and sometimes physical units. In this work, we introduce Trainspotting, a dataset of 12,068 objects and 311,886 photos from Wikimedia Commons. We reconstruct each object to recover camera poses, multi-view stereo depth, and object masks. We categorize the objects into three regimes: objects fixed at one site, objects that move between sites, and designs shared by many physical units. Evaluating recent feedforward reconstruction models on Trainspotting shows that they fail most severely in the last two regimes. Fine-tuning VGGT-Ω on Trainspotting improves camera pose, depth, and 3D geometry on held-out objects. We observe similar improvements on the independent NAVI in-the-wild benchmark, which contains photos of each object in different environments. Performance on conventional benchmarks remains largely unchanged.

The Trainspotting dataset

objects
12,068
photographs
311,886

Trainspotting is a dataset of objects reconstructed from Wikimedia Commons photos taken by many people over many years. We categorize the objects into three capture regimes: fixed objects remain at one site, roaming objects move between sites, and instance-pooled designs are photographed as different physical units.

Loading the photographic collection…

Reconstructions from Trainspotting

Our reconstruction pipeline combines dense learned matching with geometric verification. The examples below show the recovered object geometry and camera poses alongside the original photographs.

Explore all 12,068 objects ↗
Sparse reconstruction

Loading reconstruction…

Drag to orbit · shift-drag to pan

Select a camera to see the corresponding photograph.

Feedforward reconstruction with Spotter

We fine-tune VGGT-Ω on Trainspotting to obtain Spotter. Spotter outperforms existing feedforward reconstruction methods on held-out Trainspotting objects in camera pose, depth, and 3D geometry.

Input photographs Scroll to see every input →
SpotterFine-tuned on Trainspotting
Baseline reconstructionVGGT-Ω
Spotter reconstructionSpotter

Loading comparison…

Drag either view to rotate both · scroll to zoom

Evaluation on Trainspotting

1,000 held-out objects · 1,635 samples · 4 or 16 input photographs

3D geometry

% of object extent
Accuracy↓
  1. VGGT2.38
  2. π32.33
  3. Depth Anything 32.22
  4. MapAnything2.20
  5. VGGT-Ω1.84
  6. Spotter1.22
Completeness↓
  1. VGGT8.08
  2. π36.22
  3. Depth Anything 312.17
  4. MapAnything7.75
  5. VGGT-Ω12.68
  6. Spotter1.15

Depth

Relative error · inliers (%)
AbsRel↓
  1. VGGT0.160
  2. π30.162
  3. Depth Anything 30.189
  4. MapAnything0.206
  5. VGGT-Ω0.201
  6. Spotter0.071
δ1.03 (%)↑
  1. VGGT32.4
  2. π334.7
  3. Depth Anything 332.1
  4. MapAnything26.6
  5. VGGT-Ω28.1
  6. Spotter53.3

Camera pose

AUC (%)
AUC@5°↑
  1. VGGT11.5
  2. π312.5
  3. Depth Anything 310.1
  4. MapAnything5.3
  5. VGGT-Ω7.2
  6. Spotter44.2
AUC@30°↑
  1. VGGT44.1
  2. π346.4
  3. Depth Anything 329.4
  4. MapAnything33.1
  5. VGGT-Ω25.0
  6. Spotter81.2

↓ Lower is better · ↑ Higher is better

Acknowledgments

This material is based upon work supported by the Naval Air Warfare Center under Contract No. N68335-25-C-0482.

We gratefully acknowledge use of the research computing resources of the Empire AI Consortium, Inc, with support from Empire State Development of the State of New York, the Simons Foundation, and the Secunda Family Foundation.

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}
}
Download BibTeX