%0 Generic
%A Monroy Kuhn, Juan Antonio
%D 2013
%F heidok:15663
%R 10.11588/heidok.00015663
%T Morphological Analysis for Object Recognition, Matching, and Applications
%U https://archiv.ub.uni-heidelberg.de/volltextserver/15663/
%X This thesis deals with the detection and classifcation of objects in visual images and with  the analysis of shape changes between object instances. Whereas the task of object recognition  focuses on learning models which describe common properties between instances of  a specific category, the analysis of the specific differences between instances is also relevant  to understand the objects and the categories themselves. This research is governed by the  idea that important properties for the automatic perception and understanding of objects  are transmitted through their geometry or shape. Therefore, models for object recognition  and shape matching are devised which exploit the geometry and properties of the objects,  using as little user supervision as possible.  In order to learn object models for detection in a reliable manner, suitable object representations  are required. The key idea in this work is to use a richer representation of the  object shape within the object model in order to increase the description power and thus  the performance of the whole system. For this purpose, we first investigate the integration  of curvature information of shapes in the object model which is learned. Since natural  objects intrinsically exhibit curved boundaries, an object is better described if this shape  cue is integrated. This subject extends the widely used object representation based on  gradient orientation histograms by incorporating a robust histogram-based description of  curvature. We show that integrating this information substantially improves detection  results over descriptors that solely rely upon histograms of orientated gradients.  The impact of using richer shape representations for object recognition is further investigated  through a novel method which goes beyond traditional bounding-box representations  for objects. Visual recognition requires learning object models from training data. Commonly,  training samples are annotated by marking only the bounding-box of objects since  this appears to be the best trade-off between labeling information and effectiveness. However,  objects are typically not box-shaped. Thus, the usual parametrization of objects  using a bounding box seems inappropriate since such a box contains a significant amount  of background clutter. Therefore, the presented approach learns object models for detection  while simultaneously learning to segregate objects from clutter and extracting their  overall shape, without however, requiring manual segmentation of the training samples.  Shape equivalence is another interesting property related to shape. It refers to the ability  of perceiving two distinct objects as having the same or similar shape. This thesis  also explores the usage of this ability to detect objects in unsupervised scenarios, that is  where no annotation of training data is available for learning a statistical model. For this  purpose, a dataset of historical Chinese cartoons drawn during the Cultural Revolution  and immediately thereafter is analyzed. Relevant objects in this dataset are emphasized  through annuli of light rays. The idea of our method is to consider the different annuli as  shape equivalent objects, that is, as objects sharing the same shape and devise a method  to detect them. Thereafter, it is possible to indirectly infer the position, size and scale of  the emphasized objects using the annuli detections.  Not only commonalities among objects, but also the specific differences between them are  perceived by a visual system. These differences can be understood through the analysis  of how objects and their shape change. For this reason, this thesis also develops a novel  methodology for analyzing the shape deformation between a single pair of images under  missing correspondences. The key observation is that objects cannot deform arbitrarily,  but rather the deformation itself follows the geometry and constraints imposed by the  object itself. We describe the overall complex object deformation using a piecewise linear  model. Thereby, we are able to identify each of the parts in the shape which share the same deformation. Thus, we are able to understand how an object and its parts were  transformed. A remarkable property of the algorithm is the ability to automatically estimate  the model complexity according to the overall complexity of the shape deformation.  Specifically, the introduced methodology is used to analyze the deformation between original  instances and reproductions of artworks. The nature of the analyzed alterations ranges  from deliberate modifications by the artist to geometrical errors accumulated during the  reproduction process of the image. The usage of this method within this application shows  how productive the interaction between computer vision and the field of the humanities is.  The goal is not to supplant human expertise, but to enhance and deepen connoisseurship  about a given problem.