Local Feature Descriptors With Deep Hypersphere Learning

Recent works have demonstrated the power of $L_2$ normalization in local feature descriptor learning. While the descriptors are typically learned in the Euclidean space, the similarity between descriptors is often evaluated on a unit hypersphere due to the post-processing of $L_2$ normalization for descriptors, which creates a gap between the training stage and the usage stage of feature descriptors. To bridge the gap, we propose a hyperspherical descriptor learning model, where the whole network is projected onto the hyperspherical space. In addition, a squared angular triplet loss is designed to enable the proposed hyperspherical model to learn angularly discriminative descriptors. Experiments on UBC dataset show that the proposed hyperspherical descriptor outperforms its Euclidean counterparts and the state-of-the-art methods on the feature matching task.