Very Large Sets of Heuristics for Scene Interpretation

Object detection aims at automatically identifying and localizing classes of objects in still images. Software able to perform such a task is central in fields as diverse as biometric authentication, automatic surveillance or robot navigation. Most of the state-of-the-art object detection techniques are based on statistical learning. They use large sets of examples to automatically infer the regularity and specificity of a class of object to characterize it visually. The center issue experts have to deal with in such a context is invariance. They have to design adequate representations of the image based on a prior knowledge of the problem, so that the statistical learning itself can focus on unknown randomness. The standard example of such a processing is edge detection. By feeding the machine learning with an image of edges instead of the original image, one removes the need for learning invariance to illumination, which is not present in the signal anymore. Such low-level "features" are known to exist in the visual processing of animals. While fundamental, the complexity of such pre-processing steps has remained pretty low. Most of the effort has been focused on improving methods based on a restricted family of image descriptors, instead of trying to increase the versatility and richness of the feature set. The goal of this project is to investigate a new approach to object detection and machine learning in general by combining state of the art learning methods with very rich families of feature extractors. Instead of limiting ourselves to one representation of the image, we will study how to combine efficiently different families of features, and how to help experts design them. The motivation behind this project is twofold. From a practical stance we are trying to leverage the robustness resulting from the combination of a large number of modalities designed by different experts. From a more fundamental perspective, we hope to reduce the gap between artificial and biological cognition by reducing the burden on the learning part.
Machine Learning, Perceptive and Cognitive Systems
Swiss National Science Foundation
Sep 01, 2009
Aug 31, 2012