WuCongfeng
WuCongfeng
You can try to recognize the face of the original video (half body or full body), and then also recognize the face of the 512 video, and scale it to...
对齐人脸主要出现问题，1、如果1帧对1帧，2帧对2帧是肯定会抖动的，那么可以采用先对齐第一帧，后面的都对齐第一帧的坐标；但如果只对齐人脸会出现另外的问题，比如说话是，原视频是张嘴，而推理视频是闭嘴的时候，原视频的下巴就会过长，所有通常情况下，可以考虑识别人脸后扩大人脸区域做蒙版，就基本上可以解决这些问题了。
You can use the method of training the head separately, and then apply it to the half or whole body. Currently, the test results are still good, and training the...