[ACTOR] Action-Conditioned 3D Human Motion Synthesis with Transformer VAE¶

3DV Human-Motion Action-Control

Introduction¶

In this work, our goal is to take a semantic action label like “Throw” and generate an infinite number of realistic 3D human motion sequences, of varying length, that look like realistic throwing (Figure 1).

并且受使用场景限制，它需要有强约束以及相对的高效性。

涉及人体皮肤表面与主客体的交互，所以使用 SMPL 是一个非常好的选择。

noisy 3D body poses

同样采用 Positional Encoding，这与 NeRF 联系起来。

Related Work¶

• Future human motion prediction
• Human motion synthesis
• Monocular human motion estimation
• Transformer VAEs

激励

评论