References of "Teney, Damien"
Bookmark and Share    
Full Text
See detailProbabilistic Models of Visual Appearance For Object Identity, Class, and Pose Inference
Teney, Damien ULiege

Doctoral thesis (2013)

The topic of object recognition is a central challenge of computer vision. In addition to being studied as a scientific problem in its own right, it also counts many direct practical applications. We ... [more ▼]

The topic of object recognition is a central challenge of computer vision. In addition to being studied as a scientific problem in its own right, it also counts many direct practical applications. We specifically consider robotic applications involving the manipulation, and grasping of everyday objects, in the typical situations that would be encountered by personal service robots. Visual object recognition, in the large sense, is then paramount to provide a robot the sensing capabilities for scene understanding, the localization of objects of interests and the planning of actions such as the grasping of such objects. This thesis presents a number of methods that tackle the related tasks of object detection, localization, recognition, and pose estimation in 2D images, of both specific objects and of object categories. We aim at providing techniques that are the most generally applicable, by considering those different tasks as different sides of a same problem, and by not focusing on a specific type of image information or image features. We first address the use of 3D models of objects for continuous pose estimation. We represent an object by a constellation of points, corresponding to potentially observable features, which serve to define a continuous probability distribution of such features in 3D. This distribution can be projected onto the image plane, and the task of pose estimation is then to maximize its “match” with the test image. Applied to the use of edge segments as observable features, the method is capable of localizing and estimating the pose of non-textured objects, while the probabilistic formulation offers an elegant way of dealing with uncertainty in the definition of the models, which can be learned from observations — as opposed to being available as hand-made CAD models. We also propose a method, framed in a similar probabilistic formulation, in order to obtain, or reconstruct such 3D models, using multiple calibrated views of the object of interest. A larger part of this thesis is then interested in exemplar-based recognition methods, using directly 2D example images for training, without any explicit 3D information. The appearance of objects is also defined as probability distributions of observable features, defined in a nonparametric manner through kernel density estimation, using image features from multiple training examples as supporting particles. The task of object localization is cast as the cross-correlation of distributions of features of the model and of the test image, which we efficiently solve through a voting-based algorithm. We then propose several techniques to perform continuous pose estimation, yielding a precision well beyond a mere classification among the discrete, trained viewpoints. One of the proposed method in this regard consists in a generative model of appearance, capable of interpolating the appearance of learned objects (or object categories), which then allows optimizing explicitly for the pose of the object in the test image. Our model of appearance, initially defined in general terms, is applied to the use of edge segments and of intensity gradients as image features. We are particularly interested in the use of gradients extracted at a coarse scale, and defined densely across images, as they can effectively represent shape as they capture the shading onto smooth non-textured surfaces. This allows handling some cases, common in robotic applications, of objects of primitive shapes with little texture and few discriminative details, which are challenging to recognize with most existing methods. The proposed contributions, which all integrate seamlessly in a same coherent framework, proved successful on a number of tasks and datasets. Most interestingly, we obtain performance on well-studied tasks of localization in clutter and pose estimation, well above baseline methods, often on par with or superior to state-of-the-art method individually designed for each of those specific tasks, whereas the proposed framework is similarly applied to a wide range of problems. [less ▲]

Detailed reference viewed: 99 (10 ULiège)