NTH

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

AuthorsHanyu Chen, Noah Snavely

AffiliationsCornell University

November 13, 2026 2 min read
Watch on YouTube
The one-line take

Trainspotting turns scattered public photos of objects into a training resource that helps 3D reconstruction models handle real-world variation.

Key results

12068
TRAINSPOTTING objects

Objects in the Wikimedia Commons reconstruction dataset.

311886
TRAINSPOTTING photos

Photos paired with poses, depth, and object masks.

1B
VGGT-Ω parameters

Size of the pretrained model fine-tuned to create SPOTTER.

81.2
Held-out camera-pose AUC@30

SPOTTER result, up from 25.0 for VGGT-Ω.

76.0
NAVI in-the-wild camera-pose AUC@30

SPOTTER result, up from 18.8 for VGGT-Ω.

What the paper found

TRAINSPOTTING turns Wikimedia Commons archives into a benchmark for reconstructing objects photographed by different people, in different places, and across decades. It contains 12,068 objects and 311,886 photos, with camera poses, multi-view stereo depth, and object masks. Its reconstruction pipeline uses Claude Opus 5 to help curate categories, SAM3 for object gating, RoMa v2 for dense matching, COLMAP for camera recovery, and CasDiffMVS checked against MoGe-2 for depth. The dataset distinguishes fixed objects, roaming objects, and designs represented by multiple physical instances. Fine-tuning the 1B-parameter VGGT-Ω model produces SPOTTER, which raises camera-pose AUC@30 on held-out objects from 25.0 to 81.2, and on the independent NAVI in-the-wild benchmark from 18.8 to 76.0. The gains address a weakness in existing reconstruction models: difficulty aligning photos taken in changing environments or of different examples of the same design. Performance on conventional benchmarks remains largely unchanged. Training took about 12 hours on four NVIDIA B200 GPUs.

Original 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 significantly 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.

Read the original paper

More in Computer Vision

Browse all 82 papers →