Improving object detection by removing noisy samples from training sets
Gunawan Herman, Getian Ye, Jie Jeff Xu, Bang Zhang · 2008
Object detection is often formulated as a binary classification task with supervised learning that involves training datasets. Noisy samples, including mislabeled samples and ``hard-to-learn" samples, are usually found in training datasets. Such samples have a detrimental effect on the generalization performance of trained classifiers and are required to be pruned. In this paper, we propose a novel data pruning algorithm that is based on recursive Bayes approach and AdaBoost. Recursive Bayes approach increases the confidence of predictions in every iteration, while AdaBoost minimizes the number of predictions that have low confidence. Extensive experiments on real datasets show the effectiveness of the proposed algorithm in identifying and pruning noisy samples from training datasets and concurrently improving the performance of classification and object detection.