知识卡片
增强扩展的是变换覆盖
对每个训练对 (x_i,y_i) 采样变换参数 θ,生成 (T(x_i;θ),y_i),以覆盖更多允许变化。裁剪位置、翻转与颜色扰动都是可控制的参数。合成副本仍依赖原始样本,不等价于同数量独立真实观测;变换家族遗漏的语义和场景不会仅因副本多而自动出现。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
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
The way data augmentation works is that for each {x(i), y(i)} pair in your original dataset, you add M new synthetic pairs , where T is a data transformation function with stochastic parameters θ ∼ pθ. As an example, T might be the crop function, and then θ would be a set of coordinates specifying where to apply the crop. We would randomly select different
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第36章《鲁棒性与通用性训练》(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/c036.xhtml)
- 结论依据:摘录中关于增强扩展的是变换覆盖的定义、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:The way data augmentation works is that for each {x(i), y(i)} pair in your original dataset, you add M new synthetic pairs , where T is a data transformation function with stochastic parameters θ ∼ pθ. As an example, T might be the crop function, and then θ would be a set of coordinates specifying where to apply the crop. We would randomly select different