Video Question Answering Using Language-Guided Deep Compressed-Domain Video Feature
(VQAC)
This is the PyTorch Implementation of
- Nayoung Kim, Seong-Jong Ha, and Je-Won Kang. Video Question Answering Using Language-Guided Deep Compressed-Domain Video Feature. In ICCV, 2021. (to appear)
Download preprocessing data
In this experiment, we use MSVD-QA dataset. Please refer to their website for the detailed statistics of this dataset.
We already upload compressed-domain video features. You don't need to download original videos.
cd Model
Preprocessing
If you want to generate features, follow the below step. (Will be)
- Video encoding To extract motion vector and residue by HM 16.04, you need to follow this process:
- resize the video resolution: 224x224
- Feature warping