Selecting Adequate Basis Model for Adaptation in Pedestrian Detection

The emerging machine learning methods and models require large and well-annotated benchmark datasets in order to increase their generalization capacity. Pedestrian detection is one of the major research fields that suffer from data set dependency and over-fitting. A significant portion of the ongoing studies uses transfer learning or similar adaptation strategies based on deep models, which are initially trained on a benchmark dataset. This paper searches for an answer. Unfortunately, the selection of an adequate model is usually based on trial and error, which requires a tedious and time-consuming simulation through potential candidates. This paper proposes a computationally simple analysis strategy for selecting a good candidate using Shannon Entropy and Gray Level Co-occurrence Matrices, which are utilized to determine the similarity and usability of a pre-trained data set on an independent test dataset. For this purpose, a new dataset is also presented together with a detailed analysis of its characteristics. The results show that the similarity of entropy distribution of annotations (i.e. bounding boxes) has the most important characteristic to achieve robust adaptation after training,