About training data process

[cyf23]

Hi,

Thank you for your outstanding work! I have a question regarding the training process. When generating the conditioning signal, I understand that the point cloud is obtained from a single frame. However, during rendering, is the camera pose derived from running dust3r on a single frame, or from a clip of 25 frames? If it's the latter, could there be any discrepancies between the pose predicted from 25 frames during rendering and the one predicted from a single frame during inference?

Thank you for your help and for the excellent work you’ve done!

[Drexubery]

Thanks for your interest!

During training, the camera poses are derived from all 25 frames. During inference, the reference camera pose is not predicted; instead, it is fixed at (r, 0, 0) in the world coordinate system, and the subsequent camera poses are specified by the users, so there should be no discrepancies.

[HMZ76]

Thank you for your outstanding work! Could you please offer me the process training data code ?