Learning Quintuplet Loss for Large-scale Visual Geo-Localization

TL;DR

This paper introduces QUInTuplet Loss, a novel metric learning loss function designed to improve large-scale visual geo-localization by embedding multiple positive samples, addressing perspective deviations.

Contribution

The paper proposes a new QUInTuplet Loss that extends triplet loss to incorporate multiple positives, enhancing geo-localization accuracy in diverse real-world scenarios.

Findings

QUInTuplet Loss outperforms traditional triplet loss in experiments.

The method improves geo-localization accuracy across various datasets.

Enhanced robustness to perspective deviations in visual geo-localization.

Abstract

With the maturity of Artificial Intelligence (AI) technology, Large Scale Visual Geo-Localization (LSVGL) is increasingly important in urban computing, where the task is to accurately and efficiently recognize the geo-location of a given query image. The main challenge of LSVGL faced by many experiments due to the appearance of real-word places may differ in various ways. While perspective deviation almost inevitably exists between training images and query images because of the arbitrary perspective. To cope with this situation, in this paper, we in-depth analyze the limitation of triplet loss which is the most commonly used metric learning loss in state-of-the-art LSVGL framework, and propose a new QUInTuplet Loss (QUITLoss) by embedding all the potential positive samples to the primitive triplet loss. Extensive experiments have been conducted to verify the effectiveness of the proposed approach and the results demonstrate that our new loss can enhance various LSVGL methods.