Dataset Optimization for Real-Time Pedestrian Detection

Abstract

This paper tackles the problem of data selection for training set generation in the context of near real-time pedestrian detection through the introduction of a training methodology: FairTrain. After highlighting the impact of poorly chosen data on detector performance, we introduce a new data selection technique utilizing the expectation-maximization algorithm for data weighting. FairTrain also features a version of the cascade-of-rejectors enhanced with data selection principles. Experiments on the INRIA and CALTECH data sets prove that, when finely trained, a simple HoG-based detector can outperform most of its near real-time competitors.