Inflating pre-trained models

Hello All,

Upon reading the paper, I see that the authors have repeated the same image numerous time to create a "boring" video. What I fail to understand is the following, upon creating the "boring" video is the I3D model trained from scratch to create pre-trained weights? Or does there exists a schema wherein the 2-D pre-trained weights are converted to 3D in nature to initialize the network?

Start by reviewing the paper's description of the repeated-image video and the repository's I3D model context. Determine whether the pre-trained weights are learned from scratch or initialized by converting 2-D weights to 3-D, then document the evidence and clarify the training path for readers.

Written by the indexing model from the issue text.