Visual Concept Detection in Images and Videos

The rapidly increasing proliferation of digital images and videos leads to a situation where content-based search in multimedia databases becomes more and more important. A prerequisite for effective image and video search is to analyze and index media content automatically. Current approaches in the field of image and video retrieval focus on semantic concepts serving as an intermediate description to bridge the “semantic gap” between the data representation and the human interpretation. Due to the large complexity and variability in the appearance of visual concepts, the detection of arbitrary concepts represents a very challenging task. In this thesis, the following aspects of visual concept detection systems are addressed: First, enhanced local descriptors for mid-level feature coding are presented. Based on the observation that scale-invariant feature transform (SIFT) descriptors with different spatial extents yield large performance differences, a novel concept detection system is proposed that combines feature representations for different spatial extents using multiple kernel learning (MKL). A multi-modal video concept detection system is presented that relies on Bag-of-Words representations for visual and in particular for audio features. Furthermore, a method for the SIFT-based integration of color information, called color moment SIFT, is introduced. Comparative experimental results demonstrate the superior performance of the proposed systems on the Mediamill and on the VOC Challenge. Second, an approach is presented that systematically utilizes results of object detectors. Novel object-based features are generated based on object detection results using different pooling strategies. For videos, detection results are assembled to object sequences and a shot-based confidence score as well as further features, such as position, frame coverage or movement, are computed for each object class. These features are used as additional input for the support vector machine (SVM)-based concept classifiers. Thus, other related concepts can also profit from object-based features. Extensive experiments on the Mediamill, VOC and TRECVid Challenge show significant improvements in terms of retrieval performance not only for the object classes, but also in particular for a large number of indirectly related concepts. Moreover, it has been demonstrated that a few object-based features are beneficial for a large number of concept classes. On the VOC Challenge, the additional use of object-based features led to a superior performance for the image classification task of 63.8% mean average precision