Cornell's Trainspotting: 12k objects, 312k found photos to train feed-forward 3D reconstruction
kwangmoo_yi · x · 2026-11-06
Cornell researchers Hanyu Chen and Noah Snavely released Trainspotting, an object-centric 3D reconstruction dataset built from 311,886 Wikimedia Commons photos of 12,068 objects, each captured independently by many people over years. The team recovered camera poses, MVS depth, and masks, categorizing objects into fixed-site, moving, and multi-unit design regimes. Recent feed-forward models fail worst on the last two; fine-tuning VGGT-Ω on Trainspotting improves pose, depth, and geometry on held-out objects and on the NAVI in-the-wild benchmark, while conventional benchmark performance stays largely unchanged.
More from Research
- Stanford PhD JulieKallini on testing Chomsky: can LMs learn a language no human could? — ChrisGPotts · 2026-11-06
- Exo paper: agents that rewrite their own harness at runtime, cutting model cost up to 82% — sarahcat21 · 2026-11-06
- Study: LLMs Notice Context Conflicts 73% of the Time but Tell Users 0% — rseroter · 2026-11-06
- Science's Last Exam: best of 8 frontier models scores 37.5/100, all score 0 on end-to-end research — anshulkundaje · 2026-11-06
- BioSimBench: 12 frontier models tested on predicting real cell responses, best only partially right — david_van_dijk · 2026-11-06
- Hugging Face team spent $2M training MoEs — and released it all as a public handbook — Nouamanetazi · 2026-11-06