Michael Fussenegger

Generic object recognition with boosting

This paper explores the power and the limitations of weakly supervised categorization. We present a complete framework that starts with the extraction of various local regions of either discontinuity or homogeneity. A variety of local descriptors can be applied to form a set of feature vectors for each local region. Boosting is used to learn a subset of such feature vectors (weak hypotheses) and to combine them into one final hypothesis for each visual category. This combination of individual extractors and descriptors leads to recognition rates that are superior to other approaches which use only one specific extractor/descriptor setting. To explore the limitation of our system, we had to set up new, highly complex image databases that show the objects of interest at varying scales and poses, in cluttered background, and under considerable occlusion. We obtain classification results up to 81 percent ROC-equal error rate on the most complex of our databases. Our approach outperforms all comparable solutions on common databases.

Weak Hypotheses and Boosting for Generic Object Detection and Recognition

In this paper we describe the first stage of a new learning system for object detection and recognition. For our system we propose Boosting (5) as the underlying learning technique. This allows the use of very diverse sets of visual features in the learning process within a com- mon framework: Boosting — together with a weak hypotheses finder — may choose very inhomogeneous features as most relevant for combina- tion into a final hypothesis. As another advantage the weak hypotheses finder may search the weak hypotheses space without explicit calculation of all available hypotheses, reducing computation time. This contrasts the related work of Agarwal and Roth (1) where Winnow was used as learning algorithm and all weak hypotheses were calculated explicitly. In our first empirical evaluation we use four types of local descriptors: two basic ones consisting of a set of grayvalues and intensity moments and two high level descriptors: moment invariants (8) and SIFTs (12). The descriptors are calculated from local patches detected by an inter- est point operator. The weak hypotheses finder selects one of the local patches and one type of local descriptor and efficiently searches for the most discriminative similarity threshold. This differs from other work on Boosting for object recognition where simple rectangular hypotheses (22) or complex classifiers (20) have been used. In relatively simple images, where the objects are prominent, our approach yields results comparable to the state-of-the-art (3). But we also obtain very good results on more complex images, where the objects are located in arbitrary positions, poses, and scales in the images. These results indicate that our flexible approach, which also allows the inclusion of features from segmented re- gions and even spatial relationships, leads us a significant step towards generic object recognition.

A level set framework using a new incremental, robust Active Shape Model for object segmentation and tracking

Level set based approaches are widely used for image segmentation and object tracking. As these methods are usually driven by low level cues such as intensity, colour, texture, and motion they are not sufficient for many problems. To improve the segmentation and tracking results, shape priors were introduced into level set based approaches. Shape priors are generated by presenting many views a priori, but in many applications this a priori information is not available. In this paper, we present a level set based segmentation and tracking method that builds the shape model incrementally from new aspects obtained by segmentation or tracking. In addition, in order to tolerate errors during the segmentation process, we present a robust Active Shape Model, which provides a robust shape prior in each level set iteration step. For the tracking, we use a simple decision function to maintain the desired topology for multiple regions. We can even handle full occlusions and objects, which are temporarily hidden in containers by combining the decision function and our shape model. Our experiments demonstrate the improvement of the level set based segmentation and tracking using an Active Shape Model and the advantages of our incremental, robust method over standard approaches.

On-Line, incremental learning of a robust active shape model

Active Shape Models are commonly used to recognize and locate different aspects of known rigid objects. However, they require an off-line learning stage, such that the extension of an existing model requires a complete new re-training phase. Furthermore, learning is based on principal component analysis and requires perfect training data that is not corrupted by partial occlusions or imperfect segmentation. The contribution of this paper is twofold: First, we present a novel robust Active Shape Model that can handle corrupted shape data. Second, this model can be created on-line through the use of a robust incremental PCA algorithm. Thus, an already partially learned Active Shape Model can be used for segmentation of a new image in a level set framework and the result of this segmentation process can be used for an on-line update of the robust model. Our experimental results demonstrate the robustness and the flexibility of this new model, which is at the same time computationally much more efficient than previous ASMs using batch or iterated batch PCA.

A multiphase level set based segmentation framework with pose invariant shape priors

Level set based segmentation has been used with and without shape priors, to approach difficult segmentation problems in several application areas. This paper addresses two limitations of the classical level set based segmentation approaches: They usually deliver just two regions – one foreground and one background region, and if some prior information is available, they are able to take into account just one prior but not more. In these cases, one object of interest is reconstructed but other possible objects of interest and unfamiliar image structures are suppressed. The approach we propose in this paper can simultaneously handle an arbitrary number of regions and competing shape priors. Adding to that, it allows the integration of numerous pose invariant shape priors, while segmenting both known and unknown objects. Unfamiliar image structures are considered as separate regions. We use a region splitting to obtain the number of regions and the initialization of the required level set functions. In a second step, the energy of these level set functions is robustly minimized and similar regions are merged in a last step. All these steps are considering given shape priors. Experimental results demonstrate the method for arbitrary numbers of regions and competing shape priors.

Multiregion level set tracking with transformation invariant shape priors

Tracking of regions and object boundaries in an image sequence is a well studied problem in image processing and computer vision. So far, numerous approaches tracking different features of the objects (contours, regions or points of interest) have been presented. Most of these approaches have problems with robustness. Typical reasons are noisy images, objects with identical features or partial occlusions of the tracked features. In this paper we propose a novel level set based tracking approach, that allows robust tracking on noisy images. Our framework is able to track multiple regions in an image sequence, where a level set function is assigned to every region. For already known or learned objects, transformation invariant shape priors can be added to ensure a robust tracking even under partial occlusions. Furthermore, we introduce a simple decision function to maintain the desired topology for multiple regions. Experimental results demonstrate the method for arbitrary numbers of shape priors. The approach can even handle full occlusions and objects which are temporarily hidden in containers.

Object localization/segmentation using generic shape priors

Generally object segmentation is an ill-posed problem. Approaches that use only plain image information will often fail. To overcome these limitations, prior knowledge (like information of the object contour) can be added to the segmentation process. In this paper, we present a novel generic shape model. We use the expertise from the field of object class recognition, namely a boundary-fragment-model (BFM) as prior knowledge for our level set segmentation approach. Commonly, shape models need synthetically generated or pre-segmented training sets that are usually trained on one specific object or a small group of objects. With our new approach we are able to train shape models for whole categories, which makes the segmentation method much more flexible. Additionally we overcome the difficulty of the correct initialization and reduce the segmentation effort. Experimental results demonstrate the excellent performance of our method on different types of objects (categories)