Peter Hosten

Enhanced background subtraction using global motion compensation and mosaicing

Background subtraction is a widely used technique for video object segmentation. Its main drawback is its constraint to video from a static camera. Several proposals have been made to extend background model generation to camera movement, while few approaches can cope with many degrees of freedom in camera motion. We present a method to generate background images for unconstrained camera motion, zoom, rotation and even (weak) lens distortion. Our method is based on global motion estimation and a weighted summation of motion compensated images. The original contribution of our work is a statistical model that describes the deviation of local motion from global motion by a Rayleigh distribution. This allows to estimate background images where all regions that move different to global motion are suppressed, i.e. they are replaced by the appropriate background region from other frames. A quantitative evaluation on publicly available video-data shows the validity of our approach.

Annotation based personalized adaptation and presentation of videos for mobile applications

Personalized multimedia content which suits user preferences and the usage environment, and as a result improves the user experience, gains more importance. In this paper, we describe an architecture for personalized video adaptation and presentation for mobile applications which is guided by automatically generated annotations. By including this annotation information, more intelligent adaptation techniques can be realized which primarily reduce the quality of unimportant regions in case a bit rate reduction is necessary. Furthermore, a presentation layer is added to enable advanced multimedia viewers to adequately present the interesting parts of a video in case the user wants to zoom in. This architecture is the result of collaborative research done in the EU FP6 IST INTERMEDIA project.

An Evaluation of Local Features for Face Detection and Localization

Local features have the ability to overcome the major drawback of traditional, holistic object detection approaches, because they are inherently invariant to geometric deformation and pose; in addition scale and rotation invariance can be easily achieved as well. However, the selection of discriminative feature locations and local descriptions is a complex task that has not been generally solved. In case of face detection, features must possess the discriminative power to differentiate between facial parts and cluttered backgrounds while they have to remain person agnostic. A multitude of suggestions for selecting facial features for tracking or identification / recognition can be found in literature, most of which rely on semi-automatic or manual definition of the feature locations. In contrast, fully automatic feature selection and generic description approaches like SIFT and SURF have been shown to provide excellent performance for rigid as well as non-rigid registration and even for object class recognition. While quantitative evaluations exist that give a hint on the registration performance of the competing designs, these scenarios are not directly transferable to object detection. In this paper we provide qualitative and quantitative analysis of existing interest point detectors as well as local descriptions in the context of face detection and localization.

Detection of false feature correspondences in feature based object detection systems

In this paper we present a method for the detection of wrong feature correspondences in a local feature based object detection system. Common visual objects in different images share not only similar local features but also a similar spatial layout of their features. We will utilize this fact in order to distinguish between correct and wrong feature correspondences. The spatial feature layout will be modeled through a Delaunay triangulation. This triangulation is used to find clusters of feature correspondences that follow the same affine transformation. The decision whether a correspondence is correct or wrong can than be made based on this clustering. Our method is independent from the number of common objects in the images and produces reliable results even in difficult scenarios. It can also be used if the number of wrong correspondences is much higher than the number of correct correspondences. Experiments on real and synthetically generated images demonstrate the good performance of our approach.

Content-adaptive encoder optimization of the H.264/AVC deblocking filter for visual quality improvement

This paper presents a new self-adapting and content-sensitive optimization technique for the H.264/AVC in-loop deblocking filter [1], focussing on the visual enhancement of the perceived reconstruction quality. Performed frame-wisely at the encoder side, the proposed algorithm first identifies visually important image regions in the currently decoded and blocky frame, including natural and artificial edge areas, and then optimizes the filter inherent threshold decisions, finding the best trade-off between preserving natural and suppressing artificial edges. As fast optimization criterion, the low complex Edge-PSNR [2] is employed, which has proved a very good congruence to the human visual quality sensation, much better than standard PSNR. In this work, the optimization behavior of all thresholds is analyzed, showing their mostly good-natured curve shape for convex optimization, but high content dependency of the optimal values. The in-loop coding results for AVC evidence the approachs high capability for visual improvement, whose general design is easily portable to other AVC deblocking based architectures like HEVC.

Video object tracking using graph cuts and location-dependent appearance models

In this paper we present a generic video object tracking scheme that combines a feature point tracker with a graph cut based image segmentation algorithm. The system deals with deformable objects in dynamic scenes and is initialized by a ground truth of the object region in the first frame. Hence colour models representing the appearance of the foreground object and background can be estimated. Starting from a holistic approach, we extend these models by the incorporation of spatial information resulting in a section-based background and a part-based object model. The displacement vectors obtained by the feature point tracker allow for the propagation of these location-dependent appearance models into the next frame. Subsequently a graph-cut based image segmentation algorithm segments the current frame based on these models. That way a gradual tracking and segmentation of the object can be performed. Quantitative and qualitative findings demonstrate the validity of this approach.

Detection of Facial Features in Scale-Space

This paper presents a new approach to the detection of facial features. A scale adapted Harris Corner detector is used to find interest points in scale-space. These points are described by the SIFT descriptor. Thus invariance with respect to image scale, rotation and illumination is obtained. Applying a Karhunen-Loeve transform reduces the dimensionality of the feature space. In the training process these features are clustered by the k-means algorithm, followed by a cluster analysis to find the most distinctive clusters, which represent facial features in feature space. Finally, a classifier based on the nearest neighbor approach is used to decide whether the features obtained from the interest points are facial features or not.