Dynamicity and Durability in Scalable Visual Instance Search

Visual instance search involves retrieving from a collection of images the ones that contain an instance of a visual query. Systems designed for visual instance search face the major challenge of scalability: a collection of a few million images used for instance search typically creates a few billion features that must be indexed. Furthermore, as real image collections grow rapidly, systems must also provide dynamicity, i.e., be able to handle on-line insertions while concurrently serving retrieval operations. Durability, which is the ability to recover correctly from software and hardware crashes, is the natural complement of dynamicity. Durability, however, has rarely been integrated within scalable and dynamic high-dimensional indexing solutions. This article addresses the issue of dynamicity and durability for scalable indexing of very large and rapidly growing collections of local features for instance retrieval. By extending the NV-tree, a scalable disk-based high-dimensional index, we show how to implement the ACID properties of transactions which ensure both dynamicity and durability. We present a detailed performance evaluation of the transactional NV-tree: (i) We show that the insertion throughput is excellent despite the overhead for enforcing the ACID properties; (ii) We also show that this transactional index is truly scalable using a standard image benchmark embedded in collections of up to 28.5 billion high-dimensional vectors; the largest single-server evaluations reported in the literature.

💡 Research Summary

This paper addresses the critical systems challenges of scalability, dynamicity, and durability in large-scale visual instance search. Visual instance search, which aims to find specific objects (instances) within a vast collection of images, relies on indexing billions of high-dimensional local feature vectors (e.g., SIFT). As real-world image collections grow continuously, systems must support online insertions (dynamicity) while serving queries and must recover correctly from failures (durability), without sacrificing scalability.

The authors propose a solution by extending the NV-tree, a disk-based, scalable high-dimensional index. The NV-tree works by projecting data vectors onto multiple random lines and indexing the rank of each data point along each line using B+-trees. At query time, the ranks of the query vector are used to aggregate and identify approximate nearest neighbors, guaranteeing a bounded number of disk reads. The key contribution of this paper is the integration of transactional processing into the NV-tree, ensuring ACID properties (Atomicity, Consistency, Isolation, Durability) for index updates. This transformation makes the NV-tree a fully transactional dynamic index, where insertions and deletions are processed as atomic units, and the index can be recovered to a consistent state after a system crash using logging mechanisms.

The performance evaluation demonstrates the effectiveness of this transactional NV-tree in two major aspects. First, the insertion throughput remains excellent despite the overhead of enforcing ACID properties. When the index fits in memory, it achieves peak performance, and even in disk-bound scenarios, each insertion requires only a small fraction of a disk write on average. Second, the authors prove the true scalability of their system by embedding a standard image benchmark (Copydays) within distracter collections of up to 28.5 billion high-dimensional vectors—the largest single-server evaluation reported in the literature. The results show that the transactional NV-tree maintains effective retrieval quality at this unprecedented scale.

The technology described is not merely academic; it is already deployed in industrial settings, notably by Videntifier Technologies for forensic applications, where it indexes hundreds of thousands of hours of video material with daily dynamic updates. This work successfully bridges the gap between advanced computer vision needs and fundamental database system principles, delivering a practical, robust, and scalable indexing solution suitable for mission-critical, growing multimedia archives.