This thesis concerns models for visual object classes that exhibit a reasonable amount of regularity,udsuch as faces, pedestrians, cells and human brains. Such models are useful for makingud“within-object” inferences such as determining their individual characteristics and establishingudtheir identity. For example, the model could be used to predict the identity of a face, the poseudof a pedestrian or the phenotype of a cell and segment parts of a human brain.udExisting object modelling techniques have several limitations. First, most current methodsudhave targeted the above tasks individually using object specific representations; therefore, theyudcannot be applied to other problems without major alterations. Second, most methods have beenuddesigned to work with small databases which do not contain the variations in pose, illumination,udocclusion and background clutter seen in ‘real world’ images. Consequently, many existingudalgorithms fail when tested on unconstrained databases. Finally, the complexity of the trainingudprocedure in these methods makes it impractical to use large datasets.udIn this thesis, we investigate patch-based models for object classes. Our models are capableudof exploiting very large databases of objects captured in uncontrolled environments. Weudrepresent the test image with a regular grid of patches from a library of images of the sameudobject. All the domain specific information is held in this library: we use one set of images ofudthe object to help draw inferences about others. In each experimental chapter we investigateuda different within-object inference task. In particular we develop models for classification, regression,udsemantic segmentation and identity recognition. In each task, we achieve results thatudare comparable to or better than the state of the art. We conclude that patch-based representationudcan be successfully used for the above tasks and shows promise for other applications suchudas generation and localization.
展开▼
