Lequan Yu

It is not 3000fps, about 200fps.

1. I use the face detector in opencv to acquire the face bounding box. 2. It cannot reach 3000 FPS in the release mode.

I think you do not modify some paths in some cpp file. I hard-code some path name in cpp file.

我们需要得到像素差在每一个树的编码（编码由0和1组成，由于一个像素差只能落到一个叶子节点，所以这个树的编码只有一个1其余的都是0），因此我只需要知道这个像素差会落到那个叶子节点。代码中的方法可以快速得到这个叶子节点的编号。最后我只需要把feature中这个叶子节点对应的位置置为1即可。

The landmark is the mean landmark calculated from the training dataset. > 在 2016年7月17日，18:10，liuzhi136 [email protected] 写道： > > @yulequan I have read your project. It's impressive. But I'm confused that...

Hi, @kkirtac ,@muyulin. In the segmentation task, we don't do any resampling operations. For each training image, we first find the bounding box from the annotation ground truth. If the...

1. When using the sliding window, there is overlap between different windows. If the dimension is not multiple of 480, we adjust the overlap of the last sliding window. 2....

I forget the specific overlap ration. Yes, we use the simple averaging of the probabilities.

Hi, I do not have the same questions. After some research, it seems that changing the cuda and cudnn version would be a possible solution. Adding some parameters would also...

Hi, after some simple search, this problem may be related to the version of Cmake and ABI. See this link [https://github.com/lmb-freiburg/demon/issues/26]. Maybe the simple way is downgrading the TF version.