PanoImager: Geometry-Guided Novel View Synthesis and Reconstruction from Sparse Panoramic Views

Researchers have introduced two new frameworks, PanoImager and CanonicalGS, for 3D reconstruction and novel view synthesis from sparse views.

PanoImager, presented in a paper on arXiv^[1], is an SfM-free framework that reconstructs 3D scenes from sparse panoramic views. It combines feed-forward pose/depth priors, geometry-conditioned diffusion view completion, and depth-guided 3DGS optimization to improve cross-view consistency. The framework decomposes panoramic images into local perspective views and synthesizes auxiliary observations to enrich sparse evidence. Experiments on multiple benchmarks show improved stability under extreme sparsity. Meanwhile, CanonicalGS, introduced in another arXiv paper^[2], is a feed-forward pipeline for novel view synthesis that maps cluttered multi-view observations into a stable, scene-centric representation. CanonicalGS extracts view-centric evidence from depth, semantic features, and uncertainty estimates, and aggregates it in a canonical latent world using uncertainty-aware fusion. This approach improves peak signal-to-noise ratio for synthesizing novel views by up to 2.5 dB^[2] and gains 11% in semantic segmentation accuracy.