Data-Efficient Large Scale Place Recognition with Graded Similarity Supervision

Visual place recognition (VPR) is a fundamental task of computer vision for visual localization. Existing methods are trained using image pairs that either depict the same place or not. Such a binary indication does not consider continuous relations of similarity between images of the same place taken from different positions, determined by the continuous nature of camera pose. The binary similarity induces a noisy supervision signal into the training of VPR methods, which stall in local minima and require expensive hard mining algorithms to guarantee convergence. Motivated by the fact that two images of the same place only partially share visual cues due to camera pose differences, we deploy an automatic re-annotation strategy to re-label VPR datasets. We compute graded similarity labels for image pairs based on available localization metadata. Furthermore, we propose a new Generalized Contrastive Loss (GCL) that uses graded similarity labels for training contrastive networks. We demonstrate that the use of the new labels and GCL allow to dispense from hard-pair mining, and to train image descriptors that perform better in VPR by nearest neighbor search, obtaining superior or comparable results than methods that require expensive hard-pair mining and re-ranking techniques.

1. Introduction

Visual place recognition (VPR) is an important task of computer vision, and a fundamental building block of navigation systems for autonomous vehicles [24], [48]. It is approached either with structure-based methods, namely Structure-from-Motion [36] and SLAM [26], or with image retrieval [2], [15], [20], [29], [30], [46]. The former focus on precise relative camera pose estimation [34], [35]. The latter aim at learning image descriptors for effective retrieval of similar images to a given query in a nearest search approach [28]. The goal of descriptor learning is to ensure images of the same place to be projected onto close-by points in a latent space, and images of different places to be projected onto distant points [9], [10], [21]. Contrastive [19], [30] and triplet [2], [2] [2], [23], [27] loss were used for this goal and resulted in state-of-the-art performance on several VPR benchmarks. Figure 1.

(a) A place in the city of amman. (b) An image taken 6m away is labeled as positive (same place), while (c) an image taken 25.6m away is labeled as negative (not the same place) despite sharing a lot of visual cues.