In image classification, visual separability between different object categories is highly uneven, and some categories are more difficult to distinguish than others. Such difficult categories demand more dedicated classifiers. However, existing deep convolutional neural networks (CNN) are trained as flat N-way classifiers, and few efforts have been made to leverage the hierarchical structure of categories.
In this paper, we introduce hierarchical deep CNNs (HD-CNNs) by embedding deep CNNs into a category hierarchy. An HD-CNN separates easy classes using a coarse category classifier while distinguishing difficult classes using fine category classifiers. During HD-CNN training, component-wise pretraining is followed by global finetuning with a multinomial logistic loss regularized by a coarse category consistency term.
In addition, conditional executions of fine category classifiers and layer parameter compression make HD-CNNs scalable for large-scale visual recognition. We achieve state-of-the-art results on both CIFAR100 and large-scale ImageNet 1000-class benchmark datasets. In our experiments, we build up three different HD-CNNs and they lower the top-1 error of the standard CNNs by 2.65%, 3.1% and 1.1%, respectively.
While discriminative visual element mining has been introduced before, in this paper we present an approach that requires minimal annotation in both training and test time. Given only a bounding box localization of the foreground objects, our approach automatically transforms the input images into a roughly-aligned pose space and discovers the most discriminative visual fragments for each category.
These fragments are then used to learn robust classifiers that discriminate between very similar categories under challenging conditions such as large variations in pose or habitats. The minimal required input, is a critical characteristic that enables our approach to generalize over visual domains where expert knowledge is not readily available.
Moreover, our approach takes advantage of deep networks that are targeted towards fine-grained classification.It learns mid-level representations that are specific to a category and generalize well across the category instances at the same time.
Our evaluations demonstrate that the automatically learned representation based on discriminative fragments, significantly outperforms globally extracted deep features in classification accuracy.