Two-Stage Distillation for Event Image Matching

zhenjun_zhao · x · 2026-07-14

This paper discusses unsupervised target domain adaptation for event-based feature matching in unconstrained scenes.

The core idea is a two-stage distillation framework: starting with label-agnostic pre-training, followed by self-distillation using epipolar geometry constraints to enhance cross-domain matching capabilities.

Original post →

More from Embodied

Embodied channel →