3D特征：关于HFM和HBB

1.HBB    三维绑定框

(1): 要用到HBB,定义还不太清楚,来自于 VALVE Developer Community (https://developer.valvesoftware.com/wiki/Bounding_box)的 解释:

A Bounding Box is an invisible box that defines the rough size of an entity. It's chiefly used to determine when the entity is visible and to performcheapQPhysics
collision tests (often before moving on to more expensive ones: for example, bullets are only tested againsthitboxes if they are first found to intersect with the
given entity's bounding box).

一个绑定框是比较好的方法，对于物体在空间的范围表示来说。绑定球体的方法可以直接用于（x，y，z）直接使用三维下标计算点云的方法，不过这种方法必须寻找到一种合适的数据结构，否则会面对占有极大的内存空间情况。

Rotation（绑定球体）

Bounding boxes are always aligned to the world's axes. They never rotate. This is presumably because of their use for collision detection of players: it would be impossible to turn around in a tight corridor if the four corners of your bounding box
rotated with you. Axis-aligned boxes are also far cheaper to compute.

While this isn't a huge issue for objects that are roughly square from above, it makes oblong dimensions difficult to manage without switching toVPhysics.
When VPhysics isn't an option there isn't much that can be done: a Day of Defeat: Source player who has gone prone has an oblong shape that can rotate through a full 360°, which Valve could only accommodate for by creating a 'squished' bounding box that leaves the head and feet sticking out at all times (an image of this
would be nice).

关于HBB 的论文 我最先查到的是 2009-ICCV 的论文:   Multiple  Kernels for Object Detection ,具体使用方法 详见 论文内容....

2. HFM 前景假设

(2): 关于HFM  前景假设掩模，使用假设前景的方法，用于获取全局特征

详见论文 Extracting Foreground Masks towards Object Recognition,

论文详细介绍了 提取方法 和使用过程..

类似于二维图像的检测框，不过好在三维图像特别容易分割，容易从HFM中区分出目标和背景。