Humans learn robust and efficient strategies for visual tasks through interaction with their environment. In contrast, most current computer vision systems have no such learning capabilities. Motivated by insights from psychology and neurobiology, I combine machine learning and computer vision techniques to develop algorithms for visual learning in open-ended tasks. Learning is incremental and makes only weak assumptions about the task environment. I begin

We propose an unsupervised, probabilistic method for learning visual feature hierarchies. Starting from local, low-level features computed at interest point locations, the method combines these primitives into high-level abstrac-tions. Our appearance-based learning method uses lo-cal statistical analysis between features and Expectation-Maximization (EM) to identify and code spatial correla-tions. Spatial correlation is asserted when two features tend to occur at the same relative position of each other. This learning scheme results in a graphical model that allows a probabilistic representation of a flexible visual feature hier-archy. For feature detection, evidence is propagated using Nonparametric Belief Propagation (NBP), a recent gener-alization of particle filtering. In experiments, the proposed approach demonstrates efficient learning and robust detec-tion of object models in the presence of clutter and occlu-sion and under view point changes. 1.

We present a framework for learning features for visual discrimination. The learning system is exposed to a sequence of training images. Whenever it fails to recognize a visual context adequately, new features are sought that discriminate further between the true and false classes. Features consist of hierarchical combinations of primitive features (local edge and texture characteristics) that are sampled from example images. The system continues to learn better features even after all recognition errors have been eliminated, similarly to mechanisms underlying human visual expertise. Whenever the probabilistic recognition algorithm returns any posterior class probabilities greater than zero and less than one, the system attempts to find new features that improve discrimination between the classes in question. Our experiments indicate that this procedure tends to improve classification accuracy on independent test images, while reducing the number of features used fo...

Abstract not found

Abstract. We propose an unsupervised, probabilistic method for learning visual feature hierarchies. Starting from local, low-level features computed at interest point locations, the method combines these primitives into high-level abstractions. Our appearance-based learning method uses local statistical analysis between features and Expectation-Maximization to identify and code spatial correlations. Spatial correlation is asserted when two features tend to occur at the same relative position of each other. This learning scheme results in a graphical model that constitutes a probabilistic representation of a flexible visual feature hierarchy. For feature detection, evidence is propagated using Belief Propagation. Each message is represented by a Gaussian mixture where each component represents a possible location of the feature. In experiments, the proposed approach demonstrates efficient learning and robust detection of object models in the presence of clutter and occlusion and under view point changes. 1

Most existing pattern recognition techniques are based on using a fixed set of features, hand-crafted or learned by non supervised methods, to classify the data samples. But in natural and uncontrolled environments, sometimes it can be useful to use more adaptive classifiers. We propose a learning algorithm based on a boosting scheme where features are adapted to the classification task, resulting in an incremental learning approach. As a boosting scheme, a new classifier is trained at every step, but also a new feature extraction process is performed. The features are computed taking into account the most difficult examples to classify at each step, and are not imposed heuristically. The experimental results achieved show a significative increase in the learning speed as well as in the classification performance with respect to the classic boosting algorithm.

paradigm of learnable computer vision algorithms based on the