Hierarchical regression learning for car pose estimation

Abstract

Car pose estimation is a hot topic in computer vision because of its significance in intelligent transportation system. Considering the circular label space of car pose, the main limitation of classification approach is that the natural latent continuous-changing correlation across pose labels are omitted. It is more meaningful to formulate this problem as a continuous regression problem. However, the changing of light conditions and various vehicle models make imagery feature very inconsistent which easily leads global regression methods fail. In order to improve the robustness of regression learning, we proposed two novel hierarchical frameworks both of which consist of weak classifiers and strong regressors. In the first framework, Part-Aware Target Coding (PATC), the classifiers are used to predict probabilities of presence of some visible pose-sensitive parts. The probabilities together with the low level imagery features can be used as more consistent input features to train a strong regressor. The second framework, Hierarchical Sliding Slice Regression (HSSR), is in a coarse-to-fine manner. Coarse classifiers are first used to determine the belonging target group and the target group optimised fine regressors are used to estimate viewing angles. These two frameworks are applied on the benchmarking EPFL Multi-view car dataset and both of them achieve superior performance as compared to the state-of-the-art approaches.