Feature based representations for mid- and high-level vision
Tao Hai, Feng Tang · 2008
The collection of process involved in visual perception are often perceived as a hierarchy spanning the range from low-level vision through intermediate to high-level vision. Low-level feature representation is a fundamental problem in computer vision and it provides the base for mid- and high-level applications. A good feature should be effective and efficient : capable of capturing the essential image characteristics suitable for high level application and can be computed very efficiently. This thesis explores both aspects and presents two feature based image representations: feature for object tracking and efficient subspaces for image reconstruction and recognition. The dynamic feature graph representation tries to capture the appearance, structure as well as the evolving behavior of the object being tracked. This representation incorporates both distinctive local features and their geometric relations for tracking. Each feature describes the object appearance detail and the relations among features encode the object structure. The geometric relations among features are elastic and have the flexibility to handle objects with coherent motion and certain amount of variations caused by illumination changes, view point changes and articulated motions. Different aspect of essential information for effective object tracking is modeled using this representation. Two efficient subspace representations methods are proposed that use Haar-like binary box functions to represent a single image or a set of images: the nonorthogonal binary subspace (NBS) method and the binary principal component analysis (B-PCA) algorithm. A desirable property of these box functions is that their inner product operation with an image can be computed very efficiently. NBS is spanned directly by binary box functions and can be used for image representation, fast image filtering, and many other vision applications. B-PCA is a structure subspace that inherits the merits of both NBS (fast computation) and PCA (modeling data structure information). B-PCA base vectors are obtained by a novel PCA-guided NBS method. We also show that B-PCA base vectors are nearly orthogonal to each other. As a result, in the nonorthogonal vector decomposition process, the computationally intensive pseudo-inverse projection operator can be approximated by the direct dot product without causing significant distance distortion. Experiments on real image data sets show a promising performance in image matching, reconstruction, and recognition tasks with significant speed improvement.