我有一套有名字的面部肖像图像(2k)。我有一个高清视频流,我在其中发现和跟踪他们的脸。每个人在每帧,我有说5-10相似,但不同的相框块,20x20到100x100大小。我想知道如何将我检测到的面孔与OpenCV或C++中的给定描述联系起来?
发布于 2018-10-17 10:59:16
当我在下面写这些的时候,我想起了这玩意儿,这是谷歌第一个获得“人脸识别”的结果,但我从来没有试过,所以我不能评论。
在你所描述的背景下,特征脸会在一定程度上起作用。这将是理想的,如果舞台被提高了一些,并安装在讲台上,所以大多数时候,每个人都直视它,但唉。
我只是试着列出你可以做的事情/试着用不特定的顺序来提高结果。
1. One is to pick up distinct features (eyes + nose tip) and affine transform until they match the standard that is rather simple, but you lose some information, like nose length etc. And you'll have to do the same for database pictures before you train.
2. The other is to do it yourself by using face ROI coordinates in picture and transform based on that. It's easier and now that I think about it may even be more reliable. However it doesn't correct face tilt and maybe some other things.
如果角度太大,一切都会分崩离析。
https://stackoverflow.com/questions/52850611
复制相似问题