3GeM Pooling: Direction-aware and Compact Global Descriptors for Visual Place Recognition

Le résumé fourni par la source

Visual Place Recognition (VPR) is crucial for applications such as augmented reality (AR) and virtual reality (VR), where lightweight architectures and compact global descriptors enable practical deployment. Generalized Mean (GeM) pooling is widely used for global aggregation due to its simplicity and effectiveness, but its uniform spatial averaging treats all local features equally, often incorporating non-discriminative regions like pedestrians or sky. To overcome this, we propose 3GeM Pooling, which aggregates along both spatial axes and the channel dimension, capturing richer directional cues. We further introduce a Spatially Guided Channel Enhancement (SGCE) module that leverages spatially pooled information to reinforce channel-wise aggregation, producing more compact and discriminative descriptors. This design enables finer modeling of spatial structures while maintaining GeM’s lightweight nature. With equal dimensionality, 3GeM improves Recall@1 by 7.8% over GeM on MSLS-val and achieves state-of-the-art results on multiple challenging VPR benchmarks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.