Classic embedders

Before the deep-learning era, VLAD and Fisher Vector were state-of-the-art methods for image embedding. They work by extracting local descriptors from all images from the dataset, then train a clustering model on these descriptors (KMeans for VLAD, Gaussian Mixture Model for Fisher). The clustering model’s parameters are then used to compute fixed-length embeddings for the query images.

Typical local descriptors used were SIFT, RootSIFT, or SURF.

PCA is often applied to reduce the dimensionality of the local descriptors before aggregation, which can sometimes improve performance.

Table of Contents