pedestrian detection and re-identification 中的常见问题

 

有一些概念，是贯穿整个机器学习的。初初看论文的时候，根本不知所云，现在整理如下。有错的，欢迎纠正。

 

---

 

 

training、validation and testing

　　These concepts have been existed in ML for a quite while.  Both training and validation are used for parameters' adjusting.

　　We often split the training set into validation training set and validation testing set. That being said, training and validation

　　are both used for parameters' learning while testing set used for real world problems and racing.

 

---

 

 

detection methods (detectors) exploit images as input (raw images or feature maps extracted from the former) and features as detection proposals (usually generated by methods(can be a detector) ).

 

什么是detection proposal ?

 　　此坑待填

---

 

ground truth:　　在machine learning 中，ground truth 具体指的是对于监督学习的算法来说，它的training set 本身的数据的正确性。你看，如果你的训练集数据都有很大问题，那么你用它们训练出来的算法就呵呵了。在detection 里边，ground truth annotations 应该就是标记object的一个个框框。

详细请参考 wikipedia: Statistics and machine learning https://en.wikipedia.org/wiki/Ground_truth#Statistics_and_machine_learning 

 

 什么是ground truth annotation ?

　　此坑待填

---

 

IoU (Intersection-over-Union):

　　IoU，顾名思义，就是两个检测框的交的面积比上它们的并的面积。大小为0~1.0的浮点数。这个概念在tracking中很常见。

 

---

 

为什么常常说，pedestrian detection 中的问题都集中在 individual monocular colour image frames 上呢 ?

因为要在单帧图片中把人给框出来啊，为什么是 monocular 呢，因为拍的相片都是相机/摄像机啊，只有一个孔，就像用一只眼睛看的一样。

 

---