Pre-Eliminating Features for Fast Training in Real Time Object Detection in Images with a Novel Variant of AdaBoost

Our primary interest is to build fast and reliable object recognizers in images based on small training sets. This is important in cases where the training set needs to be built mostly manually, as in the case that we studied, the recognition of the Honda Accord 2004 from rear views. We described a novel variant of the AdaBoost based learning algorithm, which builds a strong classifier by incremental addition of weak classifiers (WCs) that minimize the combined error of the already selected WCs. Each WC is trained only once, and examples do not change their weights. We proposed to pre-eliminate features whose cumulative error of corresponding best WCs exceeds a predetermined threshold value. We tested two straightforward definitions of cumulative error. In both cases, we showed that, when over 97% of the initial features are eliminated at the very beginning from further training, training time is drastically reduced while having little impact on the quality of the pool of available WCs. This is a novel method that has reduced the training set WC quantity to less than 3% of its original number, greatly speeding up training time, and showing no negative impact on the quality of the final classifier. Our experiments indicated that the set of features used by Viola and Jones and others for face recognition was inefficient for our problem; therefore, each object requires its own custom-made set of features for real time and accurate recognition. Our training method, combined with the selection of appropriate features, has resulted in finding a very accurate classifier containing merely 30 weak classifiers. Compared to existing literature, we have overall achieved the design of a real time object detection machine with the least number of examples, the least number of weak classifiers, the fastest training time, and with competitive detection and false positive rates