Image Based Localization Based on Feature Scale Consistency in BOF Vector

Abstract

Image search engines commonly employ the Bag Of Features (BOF) method to represent each database image with a feature vector and retrieve the best candidate using a measure of similarity to a query image vector. The BOF vector, which specifies the occurrence frequency of features, is used with Soft Assignment (SA) to find the most similar candidates which are further analyzed using geometric information to determine the final location. In this paper, we propose a new method where partial geometric information captured in the scales of keypoints associated to feature descriptors is directly used in the feature vector entries, unlike the conventional BOF method which uses the frequency of features. The proposed method, referred to as Bag Of Scale-Indexed Features (BOSIF), is implemented with an algorithm devised to avoid the increased use of memory. A procedure for evaluating scale consistency between query and dataset images is also proposed. Experimental results demonstrate that BOSIF outperforms SA-based feature-indexed BOF method and has performance comparable to state-of-the-art approaches in terms of Recall and especially for the first retrieved images by removing false matches caused by quantization error.