Gaussian Alignment for Relative Camera Pose Estimation via Single-View Reconstruction

Yumin Li, Dylan Campbell · 2025

Read Gaussian Alignment for Relative Camera Pose Estimation via Single-View Reconstruction by Yumin Li, Dylan Campbell (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-13652.txt.

Description

Yumin Li and Dylan Campbell introduce GARPS in a 2025 computer vision paper. GARPS finds the relative pose between two cameras. It also recovers the true metric distance between them. Classic two-view methods recover direction but not scale. They also struggle with wide baselines and shiny or textureless surfaces.

GARPS skips direct image matching. It reconstructs each view separately into a three-dimensional Gaussian mixture model. A metric depth estimator and a Gaussian scene builder do that work. It then aligns the two models by optimizing a differentiable objective. The objective weighs geometry, color, spread, and semantic features. With no training of its own, GARPS beat classical and leading learned methods on the RealEstate10K dataset. The beaten methods include MASt3R.

Catalog record

CREATORYumin Li, Dylan Campbell
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-13652

Browse the full library · About the Tower of Babel