Performance analysis of object detection algorithms using small training datasets

Abstract

Object detection using machine learning approach has seen wide adoption in virtually all known industries in the past decade. Much investment and research has been put into building the most accurate object detection algorithm. However, implementation of these algorithms is only accessible to organizations with vast amount of computing and data procurement resources. In this study, the correlation of overall detection rate, training time and training sample size will be explored. In addition, threshold for minimum effective training sample size will be investigated in order to aid implementation of object detection in environments where annotated training samples are difficult to obtain. The experiment revealed that models trained using the LBP feature type performed significantly better in the 50-100 sample size range in terms of effectiveness compared to the HAAR feature type.