A bag-of-bounding-boxes approach to object-level view image retrieval

We propose a novel bag-of-words (BoW) framework to build and retrieve a compact database of view images, toward robotic localization, mapping and SLAM applications. Our method does not explain an image by many small local features (e.g. bag-of-SIFT-features) as most previous methods do. Instead, the proposed bag-of-bounding-boxes (BoBB) approach attempts to explain an image by fewer larger object patterns, which leads to a semantic and compact image descriptor. To make the view retrieval system more practical and autonomous, we do not require pre-trained object detector, but propose a novel technique for unsupervised object discovery, which is based on common pattern discovery (CPD) between the input and a known reference images. Moreover, our CPD subtask does not rely on good image segmentation techniques and is able to handle scale variations, exploiting state-of-the-art CPD techniques. Following traditional bounding box -based object annotation and knowledge transfer, we compactly describe an image in a form of bag-of-bounding-boxes (BoBB). With a slightly modified inverted file system, we efficiently index/search the BoBB descriptors. Experiments with publicly available “RobotCar” dataset show that the proposed method achieves accurate object-level view image retrieval with significantly compact image descriptors, e.g. 20 words per image.