Figure1. OurPoInit-of-Viewmethodinjectsimperceptiblepertur-
bations into input views (top) to induce cross-view inconsistency
(seedetailsinFigure3)inthefoundationalStructure-from-Motion
(SfM)initializationratherthanthesingle-view-basedlossinexist-
ing attacks. The corrupted feature correspondences propagate to
inaccurate point triangulation, leading to erroneous camera poses
and severely degraded 3D reconstruction results (bottom). Com-
paredtocleanviews,ourpoisonedinputsyieldfarfewerregistered
images, and as a result, lower quality of rendered views.