Details on pretrained models for downstream tasks (K400).

OpenGVLab / VideoMamba

[ECCV2024] VideoMamba: State Space Model for Efficient Video Understanding

https://arxiv.org/abs/2403.06977

Apache License 2.0

842 stars 60 forks source link

Details on pretrained models for downstream tasks (K400). #90

Open LSanghyeok opened 2 months ago

LSanghyeok commented 2 months ago

Hi! It's an intriguing work! I have a question about K400 pre-trained models. Did you use ImageNet-1K pre-trained models to pre-train K400 datasets for downstream tasks (e.g., Breakfast, COIN, etc) or simply train K400 from scratch without image-pretrained models?

Thanks.

Andy1621 commented 2 months ago

HI! As claimed in the caption, we use ImageNet pretrained models for K400 by default. For the models marked with †, we trained it from scratch with masked modeling.