知识卡片

普通读书笔记卡 · 8978.a · 生成模型:GPT-6

密集特征聚类需要空间对齐

理解等级:中级

从预训练模型提取C×H×W特征图,将各位置C维向量聚类后可映射回区域标签。特征步幅和感受野决定每个位置对应的输入范围,低分辨率标签恢复到原图时也需明确插值方式。可视化的PCA颜色仅是展示,不自动等于聚类实际使用的全部维度。

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

      Because human supervision is expensive and limiting, a popular alternative is to use self-supervised affinity measures, just like we saw in self-supervised contrastive learning. Popular contrastive learning methods like SimCLR [79], MoCo [193], and DINO [70], train a mapping from image patches to embedding vectors such that two patches that come from the same image embed near each other and patches from different images embed

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第31章《知觉分组》(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/c031.xhtml) - 结论依据:摘录中关于密集特征聚类需要空间对齐的定义、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:Because human supervision is expensive and limiting, a popular alternative is to use self-supervised affinity measures, just like we saw in self-supervised contrastive learning. Popular contrastive learning methods like SimCLR [79], MoCo [193], and DINO [70], train a mapping from image patches to embedding vectors such that two patches that come from the same image embed near each other and patches from different images embed