Hacker News Reader

It makes sense that they had to take out all of the cuts and fades from the training data to improve results.

I’m the background section of the research paper they mention “temporal convolution layers”, can anyone explain what that is? What sort of training data is the input to represent temporal states between images that make up a video? Or does that mean something else?

View on HN

It means that instead of (only) doing convolution in spatial dimensions, it also(/instead) happens in the temporal dimension.

A good resource for the "instead" case: https://unit8.com/resources/temporal-convolutional-networks-...

The "also" case is an example of 3D convolution, an example of a paper that uses it: https://www.cv-foundation.org/openaccess/content_iccv_2015/p...

I would assume is something similar to joining multiple frames/attentions? in channel dimension and then moving values inside so convolution will have access to some channels from other video frames.

I was working on similar idea few years ago using this paper as reference and it was working extremely well for consistency also helping with flicker. https://arxiv.org/abs/1811.08383

Reply on news.ycombinator.com