Cornell's Trainspotting: 12k 3D objects reconstructed from 311k found photos

Jimantha · x · 2026-11-07

Cornell's Hanyu Chen and Noah Snavely released Trainspotting, an object-centric 3D reconstruction dataset built from found photos on Wikimedia Commons: 12,068 objects, 311,886 photographs, with an interactive explorer for browsing sparse reconstructions, camera poses and source photos.
Core insight: state-of-the-art feedforward reconstruction models are trained on single-session captures and fail on real-world multi-view data — photos of the same object taken independently by many people across years, environments, cameras, and even different physical units. The authors split objects into fixed-site, movable, and shared-design regimes; models fail worst in the latter two.
Fine-tuning VGGT-Ω on Trainspotting improves camera pose, depth and 3D geometry on held-out objects, with similar gains on the independent NAVI in-the-wild benchmark while conventional benchmarks stay flat.
The project page lets you browse all 12k objects, including the Flying Scotsman locomotive from the teaser.

More from Multimodal