search

pcl3dv / PSAvatar

76 stars 0 forks source link

Question about the paper #7

Open prokaryote-hgy opened 4 months ago

prokaryote-hgy commented 4 months ago

Hello, nice work! May I ask if my understanding is correct: In the PMSM of the paper, each frame corresponds to a tracked FLAME mesh, and an initial point cloud is obtained through sampling. Then, during the training process, the positions of these points are optimized?

Furthermore, which parameters in the method are leanable? Are there any learnable parameters concerning with the FLAME model?

pcl3dv commented 3 months ago

Hi, thanks for your attention! First, your understanding is correct! That's how the shape model works. For the learnable parameters, we train the [color, sigma, offsets] for the shape model alignment stage, and the [rotation, scale, sh, offsets, unet] for the gaussians rendering stage, and the [pose, expression] of the FLAME and the camera are also being optimized at the same time.