Aggregating local descriptors into a compact image representation

    This paper focuses on searching images from a large dataset (10 million images or more). But here we have three parts to optimize.

1. The representation of an image
2. The dimension reduction of these representation vectors.
3. The indexing algorithm.

    In this paper, it uses a descriptor called VLAD, which is derived from Bag-of-Words and Fisher Kernel, then aggregates SIFT descriptors.
    In the second part, it uses asymmetric distance computation to do approximate nearest neighbors search. Then it uses PCA to do dimension reduction.
   We mainly focus on part 1 and part 2. And here is the flow chart[1] to describe what this paper does.

     [1] http://users.auth.gr/espyromi/publications/slides/spyromitrosWIAMIS2012slides.pdf