知识卡片
Top指标只测候选命中
Top-1检查最高分类是否真类,Top-5检查真类是否在前五名,后者通常更宽松。它们不衡量定位、计数或置信校准,单标签标注也可能遗漏图中其他合理标签;不能把Top-5高当作五个输出都准确。
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
- · 目标识别
原书出处 · 1 处
- · 查看原书依据
In some benchmarks (such as ImageNet), researchers also measure the top-5 performance, which is the percentage of times that the true label is within the set of five highest scoring predictions. Top-5 is useful in settings where the labels might be ambiguous, or where multiple labels for one image might be possible. In the case of multiple labels, ideally, the test set should specify which
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第50章《目标识别》(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/c050b.xhtml)
- 结论依据:摘录中关于Top指标只测候选命中的定义、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:In some benchmarks (such as ImageNet), researchers also measure the top-5 performance, which is the percentage of times that the true label is within the set of five highest scoring predictions. Top-5 is useful in settings where the labels might be ambiguous, or where multiple labels for one image might be possible. In the case of multiple labels, ideally, the test set should specify which