编号2605.23892优先级P2类别arXiv 新增 / 项目页会议arXiv 新增 / 项目页方法视觉几何 Transformer 的 token 选择策略，对多视图/3D foundation model 效率有参考来源arXiv / OpenReview
先说结论。它和通用视觉自监督的关系在于：视觉几何 Transformer 的 token 选择策略，对多视图/3D foundation model 效率有参考。 中高相关；详见方法、贡献和实验边界。
Figureure 1 · : We accelerate visual geometry transformers via a two-stage hierarchiFigure 1: We accelerate visual geometry transformers via a two-stage hierarchical token selection scheme: inter-frame selection followed by intra-frame selection. Our training-free method scales nearlinearly with the number of input frames, substantially improving the efficiency of visual geometry transformers with a comparable acceleration ratio with LiteVGGT [72], which requires costly full model training. Overall, our method achieves a superior trade-off between efficiency and accuracy.这张图/表用于判断 Good Token Hunting 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。Figureure 2 · : Pipeline of GoToHuntFigure 2: Pipeline of GoToHunt. Token selection is performed in the K/V space prior to the global attention layers, to determine which key/value tokens each query token interacts with. Our approach follows a two-stage hierarchical design: inter-frame selection first conducts frame-level selection, while intra-frame selection subsequently discard more tokens within each selected frame.这张图概括 Good Token Hunting 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。
核心问题
它和通用视觉自监督的关系在于：视觉几何 Transformer 的 token 选择策略，对多视图/3D foundation model 效率有参考。
方法拆解
视觉几何 Transformer 的 token 选择策略，对多视图/3D foundation model 效率有参考
