Resumen de Enhancing Visual Place Inference via Image Ranking Using Fuzzy Similarity and Agreeability
We present a fuzzy approach to visual place recognition. Our approach consists in designing a hierarchical fuzzy system to leverage matching between query and a given image while taking into account agreeability between different feature extractors, and introducing a fuzzy ranking on the basis of matching and agreeability in order to permit reranking of top-ranked images. Fuzzy ranking uses fuzzy similarity and agreeability on features determined by three different CNNs. For each of them the cosine similarity between global features representing an examined image and a query image is calculated. The cosine scores are then fed to a similarity FIS. They are also fed to an agreeability FIS, which operates on linguistic variables representing the agreeability between three cosine scores. The outcomes of Mamdani fuzzy inference systems are fed to a ranking Sugeno-type FIS. The mAP scores achieved by the proposed FIS were compared with mAP scores achieved by the similarity FIS. The algorithm has been evaluated on a large dataset for visual place recognition consisting of both images with severe (unknown) blurs and sharp images with 6-DOF viewpoint variations. Experimental results demonstrate that the mAP scores achieved by the proposed algorithm are superior to results achieved by NetVLAD as well as an algorithm combining crisp outcomes of CNNs that were investigated in this work.
