Spaces:
Running
on
Zero
Running
on
Zero
File size: 1,825 Bytes
6073e55 23fdbc0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 |
# Copyright (c) 2025 Ye Liu. Licensed under the BSD-3-Clause License.
import nncore
from torch.utils.data import Dataset
from videomind.dataset.hybrid import DATASETS
from videomind.utils.parser import parse_query, parse_question
@DATASETS.register(name='longvideobench')
class LongVideoBenchDataset(Dataset):
ANNO_PATH_VALID = 'data/longvideobench/lvb_val.json'
ANNO_PATH_TEST = 'data/longvideobench/lvb_test_wo_gt.json'
VIDEO_ROOT = 'data/longvideobench/videos_3fps_480_noaudio'
@classmethod
def load_annos(self, split='valid'):
if split == 'valid':
raw_annos = nncore.load(self.ANNO_PATH_VALID)
else:
print('WARNING: Test split does not have ground truth annotations')
raw_annos = nncore.load(self.ANNO_PATH_TEST)
annos = []
for raw_anno in raw_annos:
vid = raw_anno['video_id']
if vid.startswith('@'):
vid = vid[-19:]
# videos might come from youtube or other sources
assert len(vid) in (11, 19)
anno = dict(
source='longvideobench',
data_type='multimodal',
video_path=nncore.join(self.VIDEO_ROOT, vid + '.mp4'),
query=parse_query(raw_anno['question']),
question=parse_question(raw_anno['question']),
options=raw_anno['candidates'],
task=str(raw_anno['duration_group']),
level=raw_anno['level'],
question_category=raw_anno['question_category'])
if 'correct_choice' in raw_anno:
anno['answer'] = raw_anno['candidates'][raw_anno['correct_choice']]
anno['ans'] = chr(ord('A') + raw_anno['correct_choice'])
annos.append(anno)
return annos
|