Stochastic Attraction-Repulsion Embedding for Large Scale Image Localization

Abstract

This paper tackles the problem of large-scale imagebased localization (IBL) where the spatial location of a query image is determined by finding out the most similar reference images in a large database. For solving this problem, a critical task is to learn discriminative image representation that captures informative information relevant for localization. We propose a novel representation learning method having higher location-discriminating power.It provides the following contributions: 1) we represent a place (location) as a set of exemplar images depicting the same landmarks and aim to maximize similarities among intra-place images while minimizing similarities among inter-place images; 2) we model a similarity measure as a probability distribution on L 2 -metric distances between intra-place and inter-place image representations; 3) we propose a new Stochastic Attraction and Repulsion Embedding (SARE) loss function minimizing the KL divergence between the learned and the actual probability distributions; 4) we give theoretical comparisons between SARE, triplet ranking [2] and contrastive losses [25]. It provides insights into why SARE is better by analyzing gradients.Our SARE loss is easy to implement and pluggable to any CNN. Experiments show that our proposed method improves the localization performance on standard benchmarks by a large margin. Demonstrating the broad applicability of our method, we obtained the 3 rd place out of 209 teams in the 2018 Google Landmark Retrieval Challenge [1]. Our code and model are available at https: //github.com/Liumouliu/deepIBL. arXiv:1808.08779v2 [cs.CV] 6 Aug 2019 0 5 10 15 20 25 N-Number of top retrieved database images 86 88 90 92 94 96 98 100 Recall@N(%) Gaussian-Ind. Gaussian-Joint Cauchy-Ind. Cauchy-Joint Exponential-Ind. Exponential-Joint 0 5 10 15 20 25 N-Number of top retrieved database images 94 95 96 97 98 99 Recall@N(%) Gaussian-Ind. Gaussian-Joint Cauchy-Ind. Cauchy-Joint Exponential-Ind. Exponential-Joint 0 5 10 15 20 25 N-Number of top retrieved database images 75 80 85 90 95 Recall@N(%) Gaussian-Ind. Gaussian-Joint Cauchy-Ind. Cauchy-Joint Exponential-Ind. Exponential-Joint 0 5 10 15 20 25 N-Number of top retrieved database images 76 78 80 82 84 86 88 90 92 Recall@N(%) Gaussian-Ind. Gaussian-Joint Cauchy-Ind. Cauchy-Joint Exponential-Ind. Exponential-Joint

Lending Orientation to Neural Networks for Cross-View Geo-Localization

“…Given embeddings f g and f s of the ground-view panorama and satellite image, respectively, the aim of crossview metric learning is to embed the cross-view embeddings to a same space, with metric distances (L 2 -metric) between embeddings reflect the similarity/dissimilarity between cross-view images. There are many metric learning objective functions available, e.g., triplet ranking [5], SARE [19], contrastive [24], angular [32] losses. All losses try to pull the L 2 distances between matchable cross-view embeddings, while pushing the L 2 distances among nonmatchable cross-view embeddings.…”

University-1652: A Multi-view Multi-source Benchmark for Drone-based Geo-localization

“…The contrastive loss, pulling the distance between positive pairs, could further improve the geo-localization results [16,35]. Recently, Liu et al propose Stochastic Attraction and Repulsion Embedding (SARE) loss, minimizing the KL divergence between the learned and the actual distributions [19]. Another line of works focuses on the spatial misalignment problem in the ground-to-aerial matching.…”

Classification of Genetically Identical Left and Right Irises Using a Convolutional Neural Network

“…The final layer is the softmax layer. More details about the VGG16 architecture can be found in [14,24,25].…”