Analysis of Training Set Composition in Boosting for Object Category Recognition

This project explores recognition and classification of objects in images using the popular and effective machine learning method, Boosting. Boosting provides a framework for developing robust object detection algorithms. Boosting itself involves the training of a series of increasingly discriminating simple classifiers, and then blending their outputs – further, it involves constructing a classifier as a sum of simple weak learners. By themselves, these weak learners do not contribute much to classification performance; however, the main ingredient for Boosting’s success is a method for incrementally selecting the weak learners. That said, Boosting has successfully been applied in computer vision in areas such as pedestrian and face detection. However, despite the method’s success, its performance may be heavily dependant on the quality and quantity of the data on which it is trained. In the present context, quality of training data could cover such properties as having images that cover a range of lighting conditions, poses of an object, and alignment characteristics of the object and image. Quantity of training data really means the number of positive and negative examples of the object to be classified. In this project, I propose to explore recognition and classification of objects in images using Boosting by selecting a couple of interesting, difficult object categories, and varying the training conditions for the Boosting solution. I will do this by examining different elements of quality and quantity of training data as described, and reporting the results via standard methods such as Receiver Operating Characteristic (ROC) curves to vividly describe the contributions of quality and quantity of training data to the final, Boosting solution. This work finds that increasing data set size improves correct classification rate, classifiers trained on misaligned images perform more poorly than classifiers trained on images tightly aligned to their respective object category, and suggests that an assembly of classifiers trained on component body parts of highly articulated, jointed figures may be a better strategy for their correct classification than a classifier trained on the single, large figure itself.