Research

I build interactive systems for searching and exploring large multimedia collections, and the high-dimensional indexes that keep them fast. A second thread is evaluation, meaning how to tell whether these systems really help the people using them. Recently I have started working on composed image retrieval, which builds on the interactive learning side of this work.

Exquisitor

Exquisitor is an interactive learning system for searching and exploring large multimedia collections. It began as a user relevance feedback system that scales to 100 million images at interactive speed, and today it combines relevance feedback with conversational search and foundation models.

Exquisitor regularly takes part in interactive retrieval competitions such as the Video Browser Showdown (VBS) and the Lifelog Search Challenge (LSC). These competitions also show how differently expert and novice users work with the same system. With expert users it took 2nd place at VBS 2026, and with audience members as users it won Video Retrieval for Beginners (VR4B) at CBMI 2026.

Winner · Video Retrieval for Beginners (VR4B), CBMI 2026
2nd place · Video Browser Showdown (VBS) 2026
Winner, interactive track · IViSE Workshop, CVPR 2025

multiXview

multiXview is an interactive retrieval system for multi-stream video collections, where the same events are recorded by several egocentric and exocentric cameras.

Winner · CASTLE 2025 Interactive Grand Challenge, ACM Multimedia 2025

eCP-FS

PyPI release coming soon

eCP-FS is a disk-backed approximate nearest neighbour index built on extended Cluster Pruning (eCP), for collections too large to search, or even fit, in memory. The ecpfs library is a full re-implementation of the SISAP 2025 prototype, with a Rust core, Python bindings and a CLI.

Index nodes are loaded lazily as a search visits them, results can be pulled incrementally, new items can be inserted into a built index, and the on-disk format is plain Zarr arrays that stay easy to inspect and extend.

Evaluating Multimedia Analytics Systems

This project studies how to measure whether interactive retrieval systems work for the people using them, through automated benchmarks as well as user studies. It includes analyses of the Video Browser Showdown and the Lifelog Search Challenge, and a study of how expert and novice users perform with the same system.