README.md

Preparation

Feature Extraction

1. Video features

We follow Linjie's HERO_Video_Feature_Extractor to extract video features. If you want to use your own features, please download the raw videos of each dataset.

2. Text features

All of annotation files should be located in data.

BM-DETR
└── data
    ├── charades
    │   └── train.jsonl
    │   └── test.jsonl
    ├── charades-CD
    │   └── train.jsonl
    │   └── val.jsonl
    │   └── test_iid.jsonl
    │   └── test_ood.jsonl
    ├── QVHighlights
    │   └── train.jsonl
    │   └── val.jsonl
    │   └── test_public.jsonl
    ├── tacos
    │   └── train.jsonl
    │   └── val.jsonl
    │   └── test.jsonl
    └── activitynet
        └── train.jsonl
        └── val_1.jsonl
        └── val_2.jsonl

You can extract the text features for each dataset:

bash utils/extract_text_features/extract_text_features.sh {dset_name}

It can be one of charades, charades-CD, tacos, and activitynet.

For QVHighlights, please download official features from here.

Please see the extract_text_features.sh to check the save path.