A short paper presented at the ECCV 2026 Marine Vision workshop tested four 3D reconstruction pipelines on video from the Icefin AUV beneath the Kamb Ice Stream in Antarctica. All four failed, though with distinct symptoms: one produced near-total registration failure, another gave geometrically broken output despite healthy photometric metrics, a third generated a persistent coverage artifact, and the fourth showed unstable cross-chunk scale.

The authors argue these are not four unrelated method bugs but one shared capture regime. In all cases, the camera baseline was small relative to scene depth, so the video lacked enough parallax for reliable 3D inference regardless of the reconstruction approach.

To avoid wasting hours of compute on doomed reconstructions, they propose a detection system that identifies this failure mode in minutes. The paper does not claim to fix the problem, but it offers a way to know when under-ice reconstruction results cannot be trusted.