Fast Spectral Ranking for Similarity Search
This work addresses the challenge of efficient similarity search for object retrieval in computer vision, offering a practical solution with offline costs of a few hours and competitive query times.
The paper tackles the problem of biased Euclidean nearest neighbor search for object retrieval by introducing an explicit embedding that reduces manifold search to Euclidean search followed by dot product similarity, improving state-of-the-art on datasets like Instre with small objects and achieving query times comparable to standard search at a scale of 10^5 images.
Despite the success of deep learning on representing images for particular object retrieval, recent studies show that the learned representations still lie on manifolds in a high dimensional space. This makes the Euclidean nearest neighbor search biased for this task. Exploring the manifolds online remains expensive even if a nearest neighbor graph has been computed offline. This work introduces an explicit embedding reducing manifold search to Euclidean search followed by dot product similarity search. This is equivalent to linear graph filtering of a sparse signal in the frequency domain. To speed up online search, we compute an approximate Fourier basis of the graph offline. We improve the state of art on particular object retrieval datasets including the challenging Instre dataset containing small objects. At a scale of 10^5 images, the offline cost is only a few hours, while query time is comparable to standard similarity search.