An Analysis of the Impact of a Training Dataset Expansion for Generic Object Recognition

In the field of pattern recognition, it is well known that the contents of a training dataset affects the recognition performance. Most datasets used in the field of generic object recognition have a somewhat small number of samples per category. Thus, it is reasonable to think that we can acquire higher recognition performance by expanding the dataset. Although several approaches have attempted to increase the number of samples, the amount of these increased samples is about ten thousands which can be increased more to acquire much higher recognition performance. Thus, we collect hundred thousands of images and compare the recognition performance between the dataset with original size and the dataset which is expanded with our dataset expansion method. In addition, we also analyze the impact of using different classifiers for these expanded datasets.