Figureure 2 · Overview of the proposed GMM-EVA pipelineFig. 2. Overview of the proposed GMM-EVA pipeline. 1) A cross-modal embedding model computes frame-wise query-relevance scores. 2) A Gaussian Mixture Model (GMM), fitted via the EM algorithm, decomposes the score distribution into multiple latent semantic events. 3) An event-aware allocation strategy diferentiates visual budgets, designating one high-resolution primary keyframe and multiple low-resolution secondary keyframes per event. The keyframes in the two sets are then fed into a downstream LVLM for answer generation.这张图概括 Gaussian Mixture Modeling for Event-Aware 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 1 · Comparison of the existing keyframe selection methods and our GMM-EVA.Fig. 1. Comparison of the existing keyframe selection methods and our GMM-EVA.这张图/表用于判断 Gaussian Mixture Modeling for Event-Aware 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。
