FED-PV: A Large-Scale Synthetic Frame/Event Dataset for Particle-Based Velocimetry

TL;DR

This paper introduces FED-PV, a large-scale synthetic dataset combining frame-based and event-based particle recordings to advance particle velocimetry techniques through deep learning and data fusion.

Contribution

The paper presents FED-PV, the first large-scale dual-modal dataset for particle velocimetry, enabling research in multi-modal data fusion for flow measurement.

Findings

Generated a 350GB synthetic dataset with frame and event recordings.

Facilitates development of fusion algorithms for particle velocimetry.

Addresses the lack of cross-modal datasets in PV research.

Abstract

Particle-based velocimetry (PV) is a widely used technique for non-invasive flow field measurements in fluid mechanics. Existing PV measurements typically rely on a single type of particle recording. With advancements in deep learning and information fusion, incorporating multiple different particle recordings presents a promising avenue for next-generation PV measurement techniques. However, we argue that the lack of cross-modal datasets -- combining frame-based recordings and event-based recordings -- represents a significant bottleneck in the development of fusion measurement algorithms. To address this critical gap, we developed a dual-modal data generator FED-PV to synthesize frame-based images and event-based recordings of moving particles, resulting in a 350GB dataset generated using our approach. This generator and dataset will facilitate advancements in novel PV algorithms.