Paper accepted to ICCV 2023

A paper has been accepted for publication at the International Conference on Computer Vision (ICCV 2023).
DeDrift: Robust Similarity Search under Content Drift by Dmitry Baranchuk (Yandex), Matthijs Douse, Yash Upadhyay, Zeki Yalniz
In most practical cases, large-scale vector indexes are dynamic: new data is regularly added to the index while the deprecated data is removed. Therefore, the indexing content distribution may drift over time, resulting in index degradation. This paper investigates how content drift affects existing indexing methods and devises DeDrift, a cheap and effective method that adapts large-scale indexes to the drifting distribution on-the-fly without costly full index reconstruction.