AE2VID: Event-based Video Reconstruction via Aperture Modulation

Chenxu Bai*, Boyu Li*, Peiqi Duan†, Xinyu Zhou, Hanyue Lou, Boxin Shi†

Peking University PKU-AI² Robotics Joint Lab of Embodied AI

Paper Code

Video

AE2VID uses aperture modulation to provide dense scene cues for event-based video reconstruction.

Abstract

Event-based video reconstruction usually relies on sparse motion-triggered events, which makes static or low-motion regions difficult to recover. AE2VID periodically modulates the aperture to generate dense aperture-modulation-triggered events, then fuses them with motion-triggered events for more stable and faithful reconstruction.

Pipeline

AENet predicts dense references from aperture-triggered events. MENet reconstructs video frames from motion events and dense references.

Real Dataset Results

Qualitative reconstruction results on real AMED sequences.

Real AMED bicycle sequence reconstruction result — Bicycle sequence on the real AMED dataset.

Real AMED walking sequence reconstruction result — Walking sequence on the real AMED dataset.