Retrieval of Binary Features in Image Databases: A Study

Many state-of-the art object recognition systems rely on local image features, sometimes hundreds per image, that describe the surroundings of detected interest points by a high-dimensional feature vector. To recognize objects, these systems have to match features detected in a query image against the features stored in a database containing millions or even billions of feature vectors. Hence, efficient matching is crucial for real applications. In the past, feature vectors were often real-valued, and therefore research focused on such feature representations. Present techniques, however, involve binary features to reduce memory consumption and to speed up the feature extraction stage. Matching such binary features received relatively little attention in the computer vision community. Often, either Locality Sensitive Hashing (LSH) or quantization-based techniques, that are known from real-valued features, are used. However, an in-depth evaluation of the involved parameters in binary space has, to the best of our knowledge, not yet been performed. In this paper, we aim at closing this research gap, providing valuable insights for application-oriented research.