Compact and robust fisher descriptors for large-scale image retrieval

Vector of locally aggregated descriptors (VLAD) has overcome the lossy quantization of bag-of-words model (BOW), but its dimensionality is high for direct use. We reduce the dimensionality of VLAD by a special coding scheme. First descriptors are clustered, and then linear discriminant analysis (LDA) is performed separately within each cluster. For different cluster, we allow different dimensionality but retain the same discriminant power, aiming at optimization of total dimensionality. Furthermore, we use each feature´s nearest set of cluster centers as its expression bases, which is chosen using nearest neighbor distance ratio, so that the correspondence between a feature and its nearest set is more stable. The goal of the above scheme is to adapt the feature representation to distribution of feature classes in each cluster and distribution of cluster centers in feature space. Experiments demonstrate that our approach outperforms the state-of-the-art in computational complexity, accuracy, and robustness.