Abstract
Thanks to the powerful feature representations obtainedthrough deep convolutional neural network (CNN), theperformance of object detection has recently beensubstantially boosted. Despite the remarkable success, theproblems of object rotation, within-class variability, andbetween-class similarity remain several major challenges.To address these problems, this paper proposes a novel andeffective method to learn a rotation-invariant and Fisherdiscriminative CNN (RIFD-CNN) model. This is achievedby introducing and learning a rotation-invariant layer anda Fisher discriminative layer, respectively, on the basis ofthe existing high-capacity CNN architectures. Specifically,the rotation-invariant layer is trained by imposing anexplicit regularization constraint on the objective functionthat enforces invariance on the CNN features before andafter rotating. The Fisher discriminative layer is trained byimposing the Fisher discrimination criterion on the CNNfeatures so that they have small within-class scatter butlarge between-class separation. In the experiments, wecomprehensively evaluate the proposed method for objectdetection task on a public available aerial image datasetand the PASCAL VOC 2007 dataset. State-of-the-artresults are achieved compared with the existing baselinemethods.