TheTempAccount

Results 8 comments of TheTempAccount

I have uploaded the code for the metrics.

看一下这个json文件的绝对路径是啥,然后在运行命令里替换一下。

You can use OpenPose or AlphaPose to get the body keypoints in OpenPose format. To generate the figure in the middle you need to train a rendering model such as...

Speech2Gesture的模型代码有一些不一样,暂时还没有整理。

脸部的68个关键点的定义应该是通用的,可以查一下,把多余的点去掉。我用的是https://github.com/MVIG-SJTU/AlphaPose.git检测的,把结果转成了OpenPose的格式。 模型里只使用了12个点,输出文件里多余的点是无效的,主要是为了和OpenPose的格式一致。训练Vid2Vid的时候也没有画这些无效的点。

I don't know. There are some algorithms that convert 2D key points to 3D, but the results are not very good especially on fingers. I haven't found a satisfying method...

Accidentally close this issue..T^T

1. 不太明白你的意思 2. 重新训练需要这些,生成的话构建类似sample_audio文件夹里面的数据(.wav, .textgrid)就行了 3. 模型里不需要指定说话人,这里没有动作风格的概念,通不通用看你的目的是啥