JVET-AK0255 [AHG11] Response to Call for training materials for neural network-based video coding tool development [F. Zhang, D. Bull, J. Nawala, Y. Jiang, X. Zhu, J. Sole, E. Alshina] [late]

This contribution proposes to extend NNVC training data set and use BVI -AOM instead of BVI-DVC. Compression performance improvement for AI based algorithms is expected. Copyright conditions are more flexible than for BVI-DVC.

Total of 239 sequences, 78 new compared to BVI-DVC. The licensing statement would allow JVET establishing its own storage of the data. 39 from BVI-DVC are not included.

It was commented that the licensing allows development of standards, but not deployment. This remains an open issue.

It was suggested to investigate retraining LOP4 once with adding new sequences from BVI-AOM, and once replacing BVI-DVC by BVI-AOM (which would mean that not only the last step of training would be necessary). It was agreed to establish an EE on that.

For the EE on inter prediction, it is recommended to use the BVI-DVC extended by new sequences from BVI-AOM for the comparison of performance against the Vimeo90Ktriplet based training.

CTC would need to be changed in the next meeting, provided that benefit of the new materials is shown.

It was pointed out that contributions using this material for training need to make reference to this contribution.