ECCV paper shows scene coordinate regression models leak training-scene geometry

ducha_aiki · x · 2026-09-08

Researchers from Stony Brook, Ghent and others (ECCV 2026 Long Oral) break a common assumption: Scene Coordinate Regression (SCR) methods, which encode scenes implicitly in network weights, were presumed privacy-preserving.
They introduce a query-based attack: batch-query the model with proxy images unrelated to the target scene to get dense pixel-wise 3D coordinates, filter reliable points via stability under small perturbations, and accumulate points across batches to recover scene geometry.
From the recovered 3D representation, network features can be inverted to synthesize images from arbitrary viewpoints, revealing appearance information.
Experiments on indoor and outdoor datasets show substantial portions of training environments can be reconstructed with high geometric fidelity, exposing recognizable layouts and potentially sensitive details.

More from Safety