Binary matching for high-dimensional image descriptors

High-dimensional learning-based descriptors such as Fisher vectors (FV) is effective in encoding images, yet efficient representation of facial images in the context of large-scale databases remains a challenge for face recognition. In this paper, we propose a dimensional reduction based hashing framework to binarize high-dimensional descriptors. We introduce a compact representation of FV, and show the benefit of Linear Discriminant Analysis (LDA) combined with Local-sensitive Hashing (LSH) or Iterative Quantization (ITQ). We further present a PCA+orthogonalized LDA combined with a generalized ITQ method. Our experiments show such a framework gained decent performance. We also extend our method to single sample per person case.