← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 9046.b · 生成模型:GPT-6

重建深度可作为伪监督

理解等级:中级

多视图几何可用多张图估计场景,再把深度投回单图,形成单图预测训练对。推理只用一图,不意味着训练也只用一图;遮挡、纹理不足和重建尺度会进入伪标签。应记录训练监督来源与可靠区域。

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

      Other methods consist of using stereo pairs for training [512] or multiview geometry methods that provide more accurate 3D estimates. MegaDepth [296] uses scenes captured with multiple cameras to get good 3D reconstructions using structure from motion techniques. This allows creating a dataset of images and associated depth maps that can be used to train depth from single image methods.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第43章《单图深度估计》(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/c043.xhtml) - 结论依据:摘录中关于重建深度可作为伪监督的定义、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:Other methods consist of using stereo pairs for training [512] or multiview geometry methods that provide more accurate 3D estimates. MegaDepth [296] uses scenes captured with multiple cameras to get good 3D reconstructions using structure from motion techniques. This allows creating a dataset of images and associated depth maps that can be used to train depth from single image methods.