search

boheumd / D-NeRV

The official implementation of 'Towards Scalable Neural Representation for Diverse Videos' (CVPR 2023)
https://boheumd.github.io/D-NeRV
MIT License
46 stars 1 forks source link

About keyframe save. #4

Open haoyan14 opened 1 year ago

haoyan14 commented 1 year ago

Hey hello, I really appreciate your work and still have a few questions to ask. I want to retrain D-NeRV on the ucf101 data set I made, but the extraction method is not the same as the extraction method in your paper where your video will be extacted for 2fps per second. and I would like to ask whether there will be a big difference in the final result between the way of save the keyframes directly in the png format in torch and your way of compressing the keyframes with other methods and saving them?

haoyan14 commented 1 year ago

And also, it would be of great help if the code for dataset processing could be made public. Thanks a lot.

boheumd commented 1 year ago

Hi. If you want to achieve the state-of-the-art video compression performance, applying image compression on the keyframes are important and necessary. Without compressing the keyframes, the video compression results will be much worse due to the much larger keyframe size. For the dataset processing pipeline, you can refer to https://github.com/boheumd/D-NeRV/issues/2#issue-1854417141 for more details.