← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9003.a · 生成模型:GPT-6

容易识别不等于代表性

理解等级:中级

调试初期选择容易识别的照片有助于定位程序问题,但长期依赖这类小基准会把模型选择导向过于简单的场景。将调试集与部署评估目标区分,保留遮挡、困难姿态和复杂背景等实际变化,并按场景报告表现。

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

      When building a benchmark, especially when it is small, it is easy to introduce another photographer bias which is to take pictures that “feel” easier to recognize. This might be the right thing to do at first, when debugging a system, but, if abused, has the risk of leading us in the wrong direction when choosing which approach works best.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第35章《数据偏差与分布漂移》(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/c035b.xhtml) - 结论依据:摘录中关于容易识别不等于代表性的定义、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:When building a benchmark, especially when it is small, it is easy to introduce another photographer bias which is to take pictures that “feel” easier to recognize. This might be the right thing to do at first, when debugging a system, but, if abused, has the risk of leading us in the wrong direction when choosing which approach works best.