YuanxunLu

Of course the latter one. LSTM is a kind of RNN network, and it should take sequential data as input. 240 frames equal to 4 seconds under the 60 FPS...

First, LSTM takes sequential data as input and its output is also sequential, therefore T frames input results in T frames output. Please carefully check the definition of LSTM networks....

Any parametric monocular face reconstruction method would be an alternative, like FaceScape, DECA, 3DDFA_v2, etc.

3D facial tracking is used to obtain 3D facial landmarks and head poses. Deca is an alternative option to get these parameters. In fact, any parametric 3D facial reconstruction method...

id & scale parameters are results of 3D facial tracking. To get 3D shoulder points, we first detect 2D shoulder points using LK flow and reconstruct 3D shoulder points by...

1. velocity head pose is just the delta speed, velocity_t = pose_t+1 - pose_t. It is an implicit supervision term. You can define it as you will. 2. The 3D...

These two files are 3d facial tracking results (3D landmarks, head poses, etc.). Check the inference code and replace them with your tracking results works.

There're no overlaps between the two clips. I just use one clip for training. Check issue #27.

Check the code and the citing paper.

Smoothing on the results helps.