Fast Extraction Method of High-Level Feature Using Random Forests from Imbalanced Training Data

We propose a method of using a random forest algorithm to quickly detect semantically high-level features such as specific objects. The random forest has a lower computation cost than that of the common algorithm such as a support vector machine (SVM). However, it cannot cope with training data that have a large bias in the number of negative and positive examples. We improve the conventional training algorithm to ensure sampling the data with equal probability from each class when creating bootstrap samples, which increases the classification accuracy. Experiments on the Caltech-101 dataset resulted in a recall of 64.3% and precision of 71.1%, which were comparable to those of conventional methods. The average time needed for training and for detection were reduced to one sixteenth and one twenty-seventh that of SVM, respectively.