Global relative position space based pooling for fine-grained vehicle recognition

In the field of fine-grained vehicle recognition, large intra-class variation often affects the final prediction seriously. One of the principal reasons is unstable change of absolute positions for discriminative features, which can not be completely solved by the region of interest (ROI) localization for inevitable deviations. In this paper, we propose a global relative position space based pooling (GRPSP) method to replace the absolute position information with global relative position information, by which the discriminative features can be well aligned and thereby the intra-class variation is reduced effectively. Specifically, the discriminative features containing both absolute position information and appearance information are extracted by a convolution neural network (CNN). We then transform the absolute positions into global relative positions, and estimate position probabilities of locating at discrete representatives of global relative position space. Accordingly, the joint probabilities of different kinds of features locating at different representative positions can be calculated, and constitute the pooling vector. Lastly, the Support Vector Machine (SVM) is trained for classification. Experimental results on public and self-building challenging datasets both show that our proposed method can improve various CNNs and is the state-of-the-art.