A Hierarchy of Support Vector Machines for Pattern Detection
The Journal of Machine Learning Research
| Hi-index | 0.00 |
Detecting instances from multiple object classes simultaneously in greyscale images is a challenging problem in computer vision for many reasons. In particular, one has to deal with extreme variation in presentations, due to both linear and nonlinear transformations; for example, the variation within the class of boats or handwritten digits "3" is evidently enormous and difficult to capture, even with a very large set of examples. How does one organize the computation, both offline and online, in order to account for all presentations? Traditional approaches to multi-class object detection learn class classifiers separately and apply each of them in an exhaustive search over positions and scales, which is highly inefficient. An alternative approach, previously explored in [Gav98, GMM95, FG01, AGF04] and elsewhere, is to design the online computational process itself, rather than learning probability distributions (generative modeling) or decision boundaries (predictive learning). The general idea is to construct a tree-structured representation of the space of object instantiations as well as a binary classifier for each cell in this hierarchical scaffold. This is highly efficient due to pruning groups of hypotheses simultaneously at all stages of processing. The learned hierarchy of classifiers is applied to the image to construct a set of candidate detections, the index, which is further pruned to the final list of detected object instances by resolving confusions arising during indexing. The three (related) issues addressed in this thesis are: First , how does one construct the hierarchical representation of the space of object instantiations? Second, how does one learn a binary classifier for each cell in the hierarchy? Third, how does one prune the initial index to the final detection list? Unlike existing work (e.g., [AGF04, FG01]) on coarse-to-fine object detection, which emphasizes on design and modeling, our approach is more data-driven and focused on learning. In particular, rather than manually designing the hierarchy, we induce the hierarchy automatically from a recursive decomposition of the training set. Then, we employ boosting to build cell classifiers utilizing path-dependent "negative" examples, which are more discriminating than learning against general background patches. We also explore new region-based features, which are more robust to image distortions and background clutter than the binary edge features used in previous work, and apply various techniques to learn binary classifiers based on these new scalar features, including naive Bayes, discrete AdaBoost and real AdaBoost. (Abstract shortened by UMI.)