Orientational Spatial Part Modeling for Fine-Grained Visual Categorization

Although significant success has been achieved in fine-grained visual categorization, most of existing methods require bounding boxes or part annotations for training and test, resulting in limited usability and flexibility. To conquer these limitations, we aim to automatically detect the bounding box and parts for fine-grained object classification. The bounding boxes are acquired by a transferring strategy which infers the locations of objects from a set of annotated training images. Based on the generated bounding box, we propose a multiple-layer Orientational Spatial Part (OSP) model to generate a refined description for the object. Finally, we employ the output of deep Convolutional Neural Network (dCNN) as the feature and train a linear SVM as object classifier. Extensive experiments on public benchmark datasets manifest the impressive performance of our method, i.e., classification accuracy achieves 63.9% on CUB-200-2011 and 75.6% on Aircraft, which are actually higher than many existing methods using manual annotations.