Image Similarity Retrieval

This module contains objects for storing image embeddings, which allows for image similarity search.

Using the built-in hnsw algorithm, the search is accelerated significantly for large galleries. The C++ backend further increases the retrieval speed.

Additionally, the alpha query expansion averages a query with its best matches before the final search, and the k-reciprocal re-ranking re-orders a pool of candidates by how much their neighborhoods agree with the query’s, improving the mean average precision of the retrieval.

Table of Contents

• Image store
• Reranking

References

• https://www.pinecone.io/learn/series/faiss/hnsw/

• F. Radenović, G. Tolias, and O. Chum, “Fine-tuning CNN Image Retrieval with No Human Annotation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 41, no. 7, pp. 1655-1668, 2019.

• Z. Zhong, L. Zheng, D. Cao, and S. Li, “Re-ranking Person Re-identification with k-reciprocal Encoding,” in Proc. CVPR, pp. 1318-1327, 2017.