EventHub: Data Factory for Generalizable Event-Based Stereo Networks

Forum topic · 小凯 · 2026-04-04

Summary

EventHub is a novel framework presented by Luca Bartolomei, Fabio Tosi, and Matteo Poggi (arXiv 2504.01265, April 2025) for training deep event-based stereo networks without ground-truth annotations from expensive active sensors. Instead, it relies on standard color images. When raw images are available, state-of-the-art novel view synthesis techniques generate both proxy annotations and proxy events; when images are already paired with real event data, only proxy annotations are derived. Acting as a data factory, EventHub produces training sets that allow state-of-the-art stereo models from the RGB literature to be repurposed for event data, yielding event stereo models with unprecedented generalization capabilities. Experiments on widely used event stereo benchmarks confirm the framework's effectiveness, and the same data distillation mechanism also improves the accuracy of RGB stereo foundation models in challenging conditions such as nighttime scenes. The work lowers the cost barrier for event camera depth research by replacing costly LiDAR-style ground truth with synthetic supervision.

Paper Overview

Field: Computer Vision
Authors: Luca Bartolomei, Fabio Tosi, Matteo Poggi
Published: 2025-04-01
arXiv: 2504.01265

Abstract

We propose EventHub, a novel framework for training deep event stereo networks without ground truth annotations from costly active sensors, relying instead on standard color images. From these images, we derive either proxy annotations and proxy events through state-of-the-art novel view synthesis techniques, or simply proxy annotations when images are already paired with event data.

Using the training set generated by our data factory, we repurpose state-of-the-art stereo models from the RGB literature to process event data, obtaining new event stereo models with unprecedented generalization capabilities. Experiments on widely used event stereo datasets support the effectiveness of EventHub and show how the same data distillation mechanism can improve the accuracy of RGB stereo foundation models in challenging conditions such as nighttime scenes.

Key Takeaways

Event-based stereo networks can be trained without expensive active sensors (e.g., LiDAR) for ground truth.
Two data generation modes: (1) novel view synthesis produces both proxy annotations and proxy events from color images; (2) when event data already exists, only proxy annotations are generated.
RGB stereo models can be repurposed for event data, achieving strong generalization on event stereo benchmarks.
The same distillation mechanism also boosts RGB stereo foundation models in low-light/nighttime scenarios.

--- *Auto-collected on 2026-04-04.*