Transfer Pedestrian Detector Towards View-Adaptiveness and Efficiency

Abstract

The distribution disparity is often inevitable between the pedestrian training examples and the test data from a specific application scenario, which may result in unsatisfactory detection accuracies. In this paper, we investigate how to efficiently adapt a generic boosting-style detector for a new scenario, e.g., with a distinctive capture view-angle, with only very limited examples (e.g., ~200). The basic notation is to transfer the auxiliary knowledge encoded within the well-trained detector to a new scenario. When specific to boosting-style detectors, this auxiliary prior knowledge includes the selected features and the weights for the weak classifiers. For a new scenario, these features are reused and shifted to the most discriminative positions and scales, and the weights are further adapted by covariate shift, which introduces the covariate loss. Extensive experiments on cross-view detector adaption show the encouraging detection accuracy improvements brought by our proposed algorithm with very limited new examples.