Register assisted aggregation for visual place recognition

Visual Place Recognition (VPR) refers to use computer vision to recognize the position of the current query image. Due to the significant changes in appearance caused by season, lighting, and time spans between query and database images, these differences increase the difficulty of place recognition. Previous approaches often discard irrelevant features (such as sky, roads and vehicles) as well as features that can enhance recognition accuracy (such as buildings and trees). To address this, we propose a novel feature aggregation method designed to preserve these critical features. Specifically, we introduce additional registers on top of the original image tokens to facilitate model training, enabling the extraction of both global and local features that contain discriminative place information. Once the attention weights are reallocated, these registers will be discarded. Experimental results demonstrate that our approach effectively separates unstable features from original image representation, and achieves superior performance compared to state-of-the-art methods.