Query expansion is a popular method to improve the quality of image retrieval with both conventional and CNN representations. It has been so far limited to global image similarity. This work focuses on diffusion, a mechanism that captures the image manifold in the feature space. The diffusion is carried out on descriptors of overlapping image regions rather than on a global image descriptor like in previous approaches. An efficient off-line stage allows optional reduction in the number of stored regions. In the on-line stage, the proposed handling of unseen queries in the indexing stage removes additional computation to adjust the precomputed data. We perform diffusion through a sparse linear system solver, yielding practical query times well below one second. Experimentally, we observe a significant boost in performance of image retrieval with compact CNN descriptors on standard benchmarks, especially when the query object covers only a small part of the image. Small objects have been a common failure case of CNN-based retrieval.
@article{arxiv.1611.05113,
title = {Efficient Diffusion on Region Manifolds: Recovering Small Objects with Compact CNN Representations},
author = {Ahmet Iscen and Giorgos Tolias and Yannis Avrithis and Teddy Furon and Ondrej Chum},
journal= {arXiv preprint arXiv:1611.05113},
year = {2019}
}