Learning Token-Based Representation for Image Retrieval

In image retrieval, deep local features learned in a data-driven manner have been demonstrated effective to improve retrieval performance. To realize efficient on large database, some approaches quantize with codebook and match images aggregated kernel. However, the complexity of these is non-trivial memory footprint, which limits their capability jointly perform feature learning aggregation. generate compact global representations while maintaining regional matching capability, we propose unified framework learn representation our framework, first extract using CNNs. Then, design tokenizer module aggregate them into few visual tokens, each corresponding specific pattern. This helps remove background noise, capture more discriminative regions image. Next, refinement block introduced enhance tokens self-attention cross-attention. Finally, different are concatenated representation. The whole trained end-to-end image-level labels. Extensive experiments conducted evaluate approach, outperforms state-of-the-art methods Revisited Oxford Paris datasets.