Figureure 1 · : (a) In label-free event-image feature matching, unknown spatial relaFigure 1: (a) In label-free event-image feature matching, unknown spatial relationships and absent geometric priors prevent obtaining matching labels. (b) Supervised synthesis paradigms rely on simulated data and existing matching labels, rendering them incapable of label-free fine-tuning. (c) Alignment-dependent approaches fully rely on pixel-aligned data, failing to adapt to target datasets where such hardware relations are unavailable. (d) Our framework first performs label-agnostic pretraining for robust zero-shot features. Driven by self-distillation, the model could self-evolve directly on the unlabeled target domain.这张图概括 Label-Free Target-Domain Adaptation for Unconstrained 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。Figureure 2 · : Overview of the label-agnostic pretrainingFigure 2: Overview of the label-agnostic pretraining. Knowl edge is transferred from the frozen image teacher to the event student.这张图概括 Label-Free Target-Domain Adaptation for Unconstrained 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。
