Deep Embeddings-based Place Recognition Robust to Motion Blur

In this work we present an algorithm for severe (unknown) blur detection on RGB images. On salient CNN-based regional representations we calculate local features that are then fed to calibrated classifiers in order to estimate blur intensity. We perform scene classification and show that considerable gain in classification performance can be obtained owing to information on blur presence. We calculate global descriptors of the scene that are then fed to image retrieval engine that uses blur detection, scene category and minimum spanning tree to decide if current query image is relevant or irrelevant in context of place recognition. We show that information about blur and scene category improves mean average performance. We introduce a freely available challenging dataset both for blur detection and place recognition. It contains both images with severe blurs and sharp images with 6-DOF viewpoint variations, which were recorded using a humanoid robot.