A Comparative Review of Geometry-Aware Generative Panoramic Reconstruction in Dynamic Environments

Abstract

Panoramic reconstruction has moved from estimating a single planar transformation between photographs toward learning flexible warps, recovering explicit 3D scene structure, and synthesizing content that no camera observed. This review traces that progression through three methods chosen because each represents a distinct problem formulation: a lightweight homography-and-seam pipeline with explicit moving-object masking (our earlier work), the learned parallax-tolerant warping and composition of UDIS++, and PIS3R, which reprojects a feed-forward 3D reconstruction and refines it with point-conditioned diffusion. Rather than summarizing each paper, we compare them along technical problems: geometric alignment, dynamic objects, large parallax, missing content, computational cost, and temporal consistency. Original experiments on synthetic and real images make two distinctions concrete: depth-induced misalignment versus genuine object motion, and warping observed pixels versus generating unobserved ones. Each generation of methods resolves a different component of the problem, but none jointly models static geometry, independent motion, and generative completion. We close by proposing, as an untested research hypothesis, a geometry- and motion-conditioned generative framework governed by one principle: generate only where reliable observational evidence is insufficient.