Work

PanoLoop: Yaw‐Invariant Loop Closure for Feed‐Forward Panoramic SLAM

Year: 2026
Type: article
Abstract: Feed‐forward models such as VGGT recover camera poses and dense geometry in a single pass, and recent systems extend them into SLAM front‐ends for long pinhole sequences. Panoramic video is an even more natural fit—one 360 ° view captures the full surround and needs no intrinsic calibration—yet no feed‐forward SLAM system exists for it. We show that the crux is loop closure: well handled by pinhole pipelines, but uniquely hard on panoramas. Revisits are seen from arbitrary headings, where standard place descriptors, tied to viewing direction, fail; we therefore propose YIP, whose Fourier magnitude is exactly invariant to integer token‐grid shifts and, up to discretization, approximately invariant to physical yaw. But this invariance cuts both ways: blind to heading, it also matches look‐alike yet distant places, which an anchor‐free backbone, with no sense of absolute position, aligns into false loops. The obvious defense is to verify each candidate's geometry—our multi‐signal verifier (MSV) does, rejecting many—but no per‐candidate test is perfect, and across a flood of candidates these residual errors still admit enough false loops to corrupt the trajectory. Our key finding is that the cure is not a stronger verifier but an earlier one: candidate‐stage suppression first collapses the flood, so that MSV—now judging only a handful—becomes decisive. This sparsify‐then‐verify order beats both verifying the full flood and gating loops after pose‐graph optimization, when the trajectory is already corrupted. We call the resulting method PanoLoop. On 360Loc, it turns near‐non‐functional panoramic loop closure into an effective one, improving camera tracking over a strong panoramic visual‐odometry baseline and producing globally consistent panoramic reconstructions.
Language: English
Open Access status: bronze
OpenAlex ID: W7226007696
FWCI:
Cites: 25
Cited by:
Related to: 10