QVHighlights 1fps — Preprocessed Frames
Preprocessed version of the QVHighlights dataset for temporal video grounding. Videos are extracted at 1fps, resized to 384×384 JPEG, ready for training without any video I/O at runtime.
Contents
| File | Description |
|---|---|
annotations_train.jsonl | 7445 train annotations |
annotations_val.jsonl | 1550 val annotations |
frames folder | Train frames batch 0000–1000 |
frames_000000_001000.tar | Train frames batch 0000–1000 |
frames_001000_002000.tar | Train frames batch 1000–2000 |
frames_002000_003000.tar | Train frames batch 2000–3000 |
frames_003000_004000.tar | Train frames batch 3000–4000 |
frames_004000_005000.tar | Train frames batch 4000–5000 |
frames_005000_006000.tar | Train frames batch 5000–6000 |
frames_006000_007241.tar | Train frames batch 6000–7241 |
frames_007242_007445.tar | Remaining train frames |
val_frames_000000_end.tar | All val frames |
cache_embeddings.tar | embedings (siglips2-400m) of all the videos |
All tars extract to frames/{vid}/{timestamp:08.3f}.jpg.
JSONL Format
Each line is one annotation:
{
"vid": "6IPwMWhdKpg_60.0_210.0",
"query": "A person is riding a bike.",
"duration": 150.0,
"gt_start_sec": 12.0,
"gt_end_sec": 45.0,
"n_frames": 150
}
Source
- Annotations: jwnt4/qvhighlights-50frames (https://huggingface.co/datasets/jwnt4/qvhighlights-50frames)
- Videos: ayushsdev/qvhighlights-videos (https://huggingface.co/datasets/ayushsdev/qvhighlights-videos)