Open Access

Christian X. Ries

• In the field of computer vision, algorithms for image classification, object detection, and image retrieval are among the principal research topics. Such algorithms usually benefit from object annotations in training images indicating the locations of desired objects. In this thesis, an approach for automatically determining object annotations in the form of bounding boxes is presented. The goal of the approach is to devise bounding boxes only based on binary global image labels. That is, the only given information we want to exploit is whether or not a given training image shows the desired object. In other words, we are only given a (positive) set of images containing the object and a (negative) set of images not containing the object. Our task is then to deduce the locations of the wanted object within the positive images without further knowledge. The approach presented in this thesis is a two-stage process. In the first stage, a statistical feature model is created whichIn the field of computer vision, algorithms for image classification, object detection, and image retrieval are among the principal research topics. Such algorithms usually benefit from object annotations in training images indicating the locations of desired objects. In this thesis, an approach for automatically determining object annotations in the form of bounding boxes is presented. The goal of the approach is to devise bounding boxes only based on binary global image labels. That is, the only given information we want to exploit is whether or not a given training image shows the desired object. In other words, we are only given a (positive) set of images containing the object and a (negative) set of images not containing the object. Our task is then to deduce the locations of the wanted object within the positive images without further knowledge. The approach presented in this thesis is a two-stage process. In the first stage, a statistical feature model is created which determines visual features which are likely to be indicative for the desired object. After determining such features in the form of pixel colors and gradient features, we deduce a set of positive pixels and ultimately one or multiple bounding boxes for each positive image. We experimentally show that these boxes often already exclude a considerable amount of background from positive images. In the second stage, we further improve our bounding box estimations using a machine learning algorithm which is based on linear structural Support Vector Machines. The unknown actual object bounding boxes are modeled by latent variables which are re-estimated iteratively in our learning algorithm based on the Convex-Concave Procedure (CCCP). Thus, the final output of our algorithm are new estimations for the bounding box locations. We also provide an in-depth explanation for the latent structural SVM learning algorithm. Quantitative evaluations of all components of our approach are performed on three publicly available datasets and the results are thoroughly discussed. In additional experiments we examine the usefulness of our automatically determined bounding boxes for image retrieval and classification.…