知识卡片
几何输出表示影响损失
深度、视差、点云、法向、体素和网格表达不同结构,也让误差在空间中获得不同权重。选择输出前先明确任务需要光轴深度、相对形状还是连通表面。相同训练图与网络,改表示和损失也会改变优化偏好。
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
- · 单图深度估计
原书出处 · 1 处
- · 查看原书依据
When thinking about learning based approaches, it is always good to question the representation for the input and the output. Choosing the right representation can dramatically impact the performances of the final system (even when using the same training set and the same architecture). 3D information can be represented in many different ways (i.e., disparity maps, depth maps, planes, 3D point clouds, voxels, implicit functions,
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:摘录中关于几何输出表示影响损失的定义、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:When thinking about learning based approaches, it is always good to question the representation for the input and the output. Choosing the right representation can dramatically impact the performances of the final system (even when using the same training set and the same architecture). 3D information can be represented in many different ways (i.e., disparity maps, depth maps, planes, 3D point clouds, voxels, implicit functions,