知识卡片
隐式重采样也走逆变换
已学会图像的 f_θ 可在目标网格 q 上查询去齐次后的 M^−1q,得到变换后颜色。即使表示连续,缩小仍可能混叠;书中SIREN旋转缩放例的三脚架显示了这种问题。网络在原图外的平滑噪声外推也不是真实场景补全。
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
- · 图像与几何表示
原书出处 · 1 处
- · 查看原书依据
Once the neural network, fθ(x, y), has learned to reproduce the image, we can reconstruct the original image or apply transformations to it. Image warping is then implemented by simply applying the inverse geometric transformation to the discrete coordinates of the output grid and use the functional representation of the image to get the interpolated values:
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第38章《图像与几何表示》(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/c038b.xhtml)
- 结论依据:摘录中关于隐式重采样也走逆变换的定义、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:Once the neural network, fθ(x, y), has learned to reproduce the image, we can reconstruct the original image or apply transformations to it. Image warping is then implemented by simply applying the inverse geometric transformation to the discrete coordinates of the output grid and use the functional representation of the image to get the interpolated values: