Image Retrieval for Image-Based Localization Revisited

Abstract

To reliably determine the camera pose of an image relative to a 3D point cloud of a scene, correspondences between 2D features and 3D points are needed. Recent work has demonstrated that directly matching the features against the points outperforms methods that take an intermediate image retrieval step in terms of the number of images that can be localized successfully. Yet, direct matching is inherently less scalable than retrievalbased approaches. In this paper, we therefore analyze the algorithmic factors that cause the performance gap and identify false positive votes as the main source of the gap. Based on a detailed experimental evaluation, we show that retrieval methods using a selective voting scheme are able to outperform state-of-the-art direct matching methods. We explore how both selective voting and correspondence computation can be accelerated by using a Hamming embedding of feature descriptors. Furthermore, we introduce a new dataset with challenging query images for the evaluation of image-based localization.

Graph-Based Discriminative Learning for Location Recognition

“…For both Dubrovnik and Rome, our top 10 success rate (99.5% on Dubrovnik and 99.7% on Rome) is comparable to the results of [18] (100% / 99.7%), which uses direct 3D matching, requiring much more memory and expensive nearest neighbor computations. Our performance on Aachen dataset (89.16%) also rivals that of [26], where their best result 89.97% is achieved with a relatively expensive method, while we only use the compact set of weights learned from neighborhoods. In all cases, we improve the top k accuracies over BoW retrieval techniques, resulting in a better ranking for the final step of geometric consistency check procedure.…”

“…We evaluate our algorithm on the Dubrovnik and Rome datasets [17] and the Aachen dataset [26]; these are summarized in Table 1, along with statistics over the neighborhoods we compute. To represent images as BoW histograms, we learn two kinds of visual vocabularies [22]: one vocabulary learned from each dataset itself (a specific vocabulary) and another shared vocabulary learned from ∼20,000 randomly sampled images from an unrelated dataset (a generic vocabulary).…”

“…As with other location recognition approaches [27,12,14,26], our work uses an image-retrieval-based framework using a bag-ofwords model for a database of images. However, our goal is not retrieval per se (i.e., to retrieve all related instances of a query image), but instead recognition, where we aim to determine where an image was taken (for which a single correctly retrieved database image can be sufficient).…”

Minimal Scene Descriptions from Structure from Motion Models

“…In this section, we evaluate the performance of our algorithm on several datasets, including the Dubrovnik dataset of Li et al [9], the Aachen dataset of Sattler et al [16], and the much larger Landmarks dataset [10]; these three datasets are summarized in Table 1.…”

“…We evaluate three approaches to computing minimal scene descriptions: the K-cover algorithm (KC) [9], our initial point set selection algorithm only (KCD), and our Dataset # DB Imgs # 3D Points # Queries Dubrovnik [9] 6,044 1,886,884 800 Aachen [16] 4,479 1,980,036 369 Landmarks [10] 205,813 38,190,865 10,000 full approach including the probabilistic K-cover algorithm (KCP). All methods output a list of points to keep in the original 3D point cloud database.…”