How to Train Your Dragon: Best Practices in Pedestrian Classifier Training

Abstract

The present year witnesses another milestone in Pedestrian detection's journey. It has achieved remarkable progress in the course of the past 15 years, and experts foresee an everyday use of numerous stemming applications within the next 15 years. Standing on the tipping point between yesterday and tomorrow pushes to the field's retrospect. Features, context, the combination of approaches, and feature data are responsible for most of the breakthroughs in pedestrian detection. If the first three elements cover the bulk of the literature, much fewer efforts have been dedicated to featuring data. In many aspects, the construction of a training set remains similar to what it was at the birth of the domain, and some related problems are not well studied and sometimes still tackled empirically. This paper gets down to the study of pedestrian classifier training conditions. More than a survey of existing training classifiers or features, our goal is to highlight impactful parameters, potential new research directions, and combination dilemmas. Our findings are experimentally verified on two major datasets: the INRIA and Caltech-USA datasets.

Keywords

• Pedestrian detection
• real-time
• data selection
• parameter estimation
• big data
• dataset training