Perspective-invariant imaging

Paper Title:

Perspective-Equivariant Imaging: an Unsupervised Framework for Multispectral Pansharpening

Published on:

14 March 2024

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Andrew Wang,

Mike Davies

Bullets

Key Details

•

Proposes perspective-equivariant imaging framework that leverages perspective invariance

•

Extends previous equivariant imaging work to richer non-linear transforms

•

Achieves SOTA unsupervised results on satellite pansharpening

•

Robust to noise compared to other unsupervised methods

•

Easy to train and fast inference

Explore the topics in this paper

image reconstruction

pansharpening

perspective-equivariant imaging

satellite imagery

unsupervised learning

AI generated summary

Perspective-invariant imaging

This paper proposes a new framework called perspective-equivariant imaging that uses the perspective invariance of images from camera systems to solve ill-posed image reconstruction problems without ground truth data. It is applied to multispectral satellite image pansharpening and outperforms other unsupervised methods, achieving state of the art results.