Pano‐Matching: Panoramic Local Feature Matching Using Interconnected Dynamic Transformer

Abstract

ABSTRACT With the popularity of consumer‐level panoramic cameras, panoramic projection is increasingly used in various industries, such as social communication, entertainment, and education. The utilization of image information, particularly in panoramic projection, predominantly depends on image matching. This technology involves extracting key features from images to meet the requirements of subsequent tasks. However, due to the distortion caused by geometric deformation in panoramic projection, traditional planar image matching methods encounter challenges in key point detection, matching accuracy, and pose estimation, often leading to failure or suboptimal performance. To address this challenge, a novel detector‐free matching method is proposed. Pano‐matching introduces two key innovations: a clustering‐based dynamic pruning scheme to accelerate attention convergence by focusing on valid feature pairs, and a redesigned fine‐level matching approach that effectively leverages both the central feature vector and its local neighbourhood. These innovations allow pano‐matching to handle the distortions in panoramic images and outperform existing convolutional neural network‐based methods in both accuracy and computational efficiency. Experimental results demonstrate that pano‐matching achieves state‐of‐the‐art performance in pose estimation and feature matching, significantly improving over current panoramic and planar matching methods.