Domain Adaptation with One-step Transformation

It is a crucial property for autonomous vehicle driving systems to robustly perform in different driving surroundings. However, the modules based on computer vision suffer from the performance degradation problem, when there is distribution discrepancy between the practically captured data and the training data. In this paper, we address this problem by learning an one-step transformation to bridge the discrepancy from source domain to target domain. Since the feature space learned by labeled source data is well-trained, the target data firstly are directly mapped to this feature space. With regard the domain discrepancy, the distribution of source and target features need to be further aligned. We model the aligning process as an one-step transformation and implement it as one layer convolutional neural network. In order to effectively learn the one-step transformation, a new adversarial loss function is proposed to minimize the Wasserstein distance of involving domains and the prediction error simultaneously. The experiments are conducted on six datasets, including the challenging traffic-related data,e.g. traffic sign images and the pedestrian fisheye images captured by the cameras installed in a moving vehicle. The results demonstrated the efficiency of the proposed method in comparison with other eight classical recognition methods.