← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9093 · 生成模型:GPT-6

多任务损失需区分有效对象

理解等级:中级

实例分割可联合分类、定位和掩码损失,只对有效正例计算后两项,通过权重平衡数值与目标。三项量级、监督范围和匹配方式都会影响学习;损失相加不保证各子任务自动协同或全部同等改善。

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

      The first classification loss corresponds to the classification of the crop as containing the object or not; if the object is present, the second term measures the accuracy of the bounding box, and the third term measures whether the segmentation, ŝi, inside each bounding box matches the ground truth instance segmentation, si. The parameters λ1 and λ2 control how much weight each loss has on

    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:摘录中关于多任务损失需区分有效对象的定义、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 first classification loss corresponds to the classification of the crop as containing the object or not; if the object is present, the second term measures the accuracy of the bounding box, and the third term measures whether the segmentation, ŝi, inside each bounding box matches the ground truth instance segmentation, si. The parameters λ1 and λ2 control how much weight each loss has on