Request of a best pretrained checkpoint for downstream tasks about videomae

[Lycus99]

Theoretically, the feature learned by VideoMAE is powerful for video understanding. I wonder to know that is VideoMAE pretrained model useful for other downstream tasks such as action segmentation? the authors do not give the results on Breakfast or Salad50 or GTEA

[makecent]

@Lycus99 For example the AVA dataset for spatial-temporal action deteciton.

Comparing the last two rows, pretraining on K700 with further supervised training can get better performance

[Lycus99]

thank you very much！

[makecent]

I mixed up the concept of "Extra Labels" in the paper and "Extra Data" in the Model Zoo. The checkpoint in "Fine-tune" column should already be "(self + fully) supervised" pretrained.