How H3-World Aligns 11.3 Hours of ABot Data Into Action Supervision
A review of the H3-World paper and official processing code finds that its 7,872 training windows amount to about 11.30 hours under a 124-frame, 24-FPS calculation, rather than the “500h” implied by the ABot dataset name. The pipeline recovers camera motion from COLMAP poses, aligns video and key states by source-frame index, pools them into 37 latent intervals, and ultimately represents eight keys plus a camera-speed bit as text. The audit notes that this compression can discard timing, magnitude, and ordering details, while its arithmetic examples did not constitute a full audit of the actual data or training run.
The analysis separates H3-World’s recomputable training-window budget from the broader ABot dataset name and shows that reproducing the model requires tracing how raw video, poses, and key states become compressed action text.
H3 -World 的 11.3 小时训练数据,是怎样对齐成动作监督的?
作者解释或未执行记录;人造演算不代表真实数据质量或训练运行结果。
H3-World 从名字带有“500h”的 ABot 数据集取数据,论文实验却只用了 7,872 个训练片段。按每段 124 帧、24 FPS 计算,论文声明的训练窗口按帧预算折算约为 11.30 小时。这个落差首先改变的是我们要核查的问题:复现时不能拿来源库的名称回答“训练了多少”,也不能把一段视频旁边放上按键文件,就认为已经获得可用的动作监督。
更值得看的是数据经过了什么变化。官方处理程序没有直接使用动作 JSON 中现成的连续变化数组。它从 COLMAP 相机姿态恢复运动,把视频与键态裁到共同的源帧索引,再把这些记录压到视频 latent 的时间区间里。最后给模型的并非原始姿态数值,而是由八个键与一个相机速度位生成的文本。这条链每一步都可能丢信息,检验目标也各不相同。