知识卡片

普通读书笔记卡 · 9111.a · 生成模型:GPT-6

混合数据训练需处理尺度歧义

理解等级:中级

不同深度数据可能有不同尺度、相机和立体基线,尺度与偏移不敏感的训练目标有助于混合利用。代价是输出未必直接具有统一米制意义;需要分清提高训练兼容性与保留可测量绝对量之间的取舍。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      As the model is trained with a diverse collections of data sources, there is an ambiguity on the overall scale, camera parameters, and stereo baselines. Therefore, the loss is defined in a way that it can be invariant to all those ambiguities. The invariant loss makes the training more robust, but it also makes the output less convenient to use.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第55章《重访简单视觉系统》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c055b.xhtml) - 结论依据:摘录中关于混合数据训练需处理尺度歧义的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:As the model is trained with a diverse collections of data sources, there is an ambiguity on the overall scale, camera parameters, and stereo baselines. Therefore, the loss is defined in a way that it can be invariant to all those ambiguities. The invariant loss makes the training more robust, but it also makes the output less convenient to use.