Adaptive pedestrian detection by modulating features in dynamical environment

The accuracy of a trained pedestrian detector is always decreased in a new scenario, if the distributions of the samples in the testing and training scenarios are different. Traditional methods solve this problem based on domain adaption techniques. Unfortunately, most of existing methods need to keep source samples or label target samples in the detection phase. Therefore, they are hard to be applied in the real applications with dynamical environment. For this problem, we propose a feature modulation model, which consists of a Simple Dynamical Neural Network (SDNN) and a Modulating Neural Network (MNN). In SDNN, a dynamical layer is adopt to adaptively weight the feature maps, whose parameters are predicted by MNN. For each candidate proposal, the SDNN generates a proprietary deep feature respectively. Our contributions include 1) the first feature-based unsupervised domain adaptation method which is very suitable for real applications and 2) a new scheme of dynamically weighting feature maps, in which the corresponding training method is also given. Experimental results confirm that our method can achieve the competitive results on two pedestrian datasets.