Abstract: In this paper we present the importance of including the user in the loop in a handwritten word spotting framework. Several off-the-shelf query fusion and relevance feedback strategies have been tested in the handwritten word spotting context. The increase in terms of precision when the user is included in the loop is assessed using two datasets of historical handwritten documents and a baseline word spotting approach based on a bag-of-visual-words model.

Abstract: Mode detection in online handwritten documents refers to the process of distinguishing different types of contents, such as text, formulas, diagrams, or tables, one from another. In this paper a new approach to mode detection is proposed that uses bidirectional long-short term memory (BLSTM) neural networks. The BLSTM neural network is a novel type of recursive neural network that has been successfully applied in speech and handwriting recognition. In this paper we show that it has the potential to significantly outperform traditional methods for mode detection, which are usually based on stroke classification. As a further advantage over previous approaches, the proposed system is trainable and does not rely on user-defined heuristics. Moreover, it can be easily adapted to new or additional types of modes by just providing the system with new training data.

Abstract: Clinical studies on atherosclerosis regression/progression performed by Intravascular Ultrasound analysis require the alignment of pullbacks of the same patient before and after clinical interventions. In this paper, a methodology for the automatic alignment of IVUS sequences based on the Dynamic Time Warping technique is proposed. The method is adapted to the specific IVUS alignment task by applying the non-rigid alignment technique to multidimensional morphological signals, and by introducing a sliding window approach together with a regularization term. To show the effectiveness of our method, an extensive validation is performed both on synthetic data and in-vivo IVUS sequences. The proposed method is robust to stent deployment and post dilation surgery and reaches an alignment error of approximately 0.7 mm for in-vivo data, which is comparable to the inter-observer variability.

Abstract: Error-correcting output codes (ECOC) represents a powerful framework to deal with multiclass classification problems based on combining binary classifiers. The key factor affecting the performance of ECOC methods is the independence of binary classifiers, without which the ECOC method would be ineffective. In spite of its ability on classification of problems with relatively large number of classes, it has been applied in few real world problems. In this paper, we investigate the behavior of the ECOC approach on two image vision problems: logo recognition and shape classification using Decision Tree and AdaBoost as the base learners. The results show that the ECOC method can be used to improve the classification performance in comparison with the classical multiclass approaches.

Abstract: Poster
paper TA.P5.12
Face detectors typically produce a large number of false positives and this leads to the need to have a further non maximum suppression stage to eliminate multiple and spurious responses. This stage is based on considering spatial heuristics: true positive responses are selected by implicitly considering several restrictions on the spatial distribution of detector responses in natural images. In this paper we analyze the limitations of this approach and propose an efficient search method to overcome them. Results show how the application of this new non-maximum suppression approach to a simple face detector boosts its performance to state of the art results.

Abstract: This paper addresses the problem of video surveillance by mobile cameras. We present a method that allows online change detection in night-time outdoor surveillance. Because of the camera movement, background frames are not available and must be “localized” in former sequences and registered with the current frames. To this end, we propose a Frame Localization And Registration (FLAR) approach that solves the problem efficiently. Frames of former sequences define a database which is queried by current frames in turn. To quickly retrieve nearest neighbors, database is indexed through a visual dictionary method based on the SURF descriptor. Furthermore, the frame localization is benefited by a temporal filter that exploits the temporal coherence of videos. Next, the recently proposed ECC alignment scheme is used to spatially register the synchronized frames. Finally, change detection methods apply to aligned frames in order to mark suspicious areas. Experiments with real night sequences recorded by in-vehicle cameras demonstrate the performance of the proposed method and verify its efficiency and effectiveness against other methods.

Abstract: Vision-based object detectors are crucial for different applications. They rely on learnt object models. Ideally, we would like to deploy our vision system in the scenario where it must operate, and lead it to self-learn how to distinguish the objects of interest, i.e., without human intervention. However, the learning of each object model requires labelled samples collected through a tiresome manual process. For instance, we are interested in exploring the self-training of a pedestrian detector for driver assistance systems. Our first approach to avoid manual labelling consisted in the use of samples coming from realistic computer graphics, so that their labels are automatically available [12]. This would make possible the desired self-training of our pedestrian detector. However, as we showed in [14], between virtual and real worlds it may be a dataset shift. In order to overcome it, we propose the use of unsupervised domain adaptation techniques that avoid human intervention during the adaptation process. In particular, this paper explores the use of the transductive SVM (T-SVM) learning algorithm in order to adapt virtual and real worlds for pedestrian detection (Fig. 1).