yeates / dmt Goto Github PK

WIll this be able to efficiently process longer videos?

From the paper also, in the 4.2 ablation study, I don't find any clear element why the network need to be trained in two stages (DMTimg and DMTvid) especially cause you have claimed:

Therefore, we train DMTvid from scratch instead of fine-tuning it on DMTimg

Was there any specific issue to unify the loss to a have an E2E single stage approach?

In 6.6 section of the paper

Firstly, while our method exhibits ro-
bustness to random video shapes, one common drawback of
Transformers-based methods [25, 53, 23] is their high mem-
ory requirements when processing high-resolution content,
such as 1080p videos.

Do you have any memory requirement to share at specific input resolution?

I have been waiting for the demo and code. Is there an ETA yet for it?

Dear Authors,
Awesome idea!
When will you be able to release your code for us to try?
Thanks

Great works！ I've been reading your paper and am interested in the Receptive Field Contextualizer (RFC) module mentioned in Section 3.2. Could you provide more details about the RFC's structure, especially why a kernel size of K = 13 was chosen? Your insights would be greatly appreciated!