Something-Something V2
Something-Something V2 is a video action-recognition dataset of short clips showing people performing fine-grained actions with everyday objects.
The dataset is distributed by Qualcomm under a non-standard data license. The
builder uses the Qualcomm download package and follows the normal
stable-datasets cache rules. On machines with small home quotas, set
STABLE_DATASETS_CACHE_DIR before loading the dataset, or pass
download_dir= and processed_cache_dir= explicitly.
Usage
from stable_datasets.video import SomethingSomethingV2
ds = SomethingSomethingV2(split="train")
sample = ds[0]
video = sample["video"] # VideoRef
label = sample["label"] # int class id, or -1 if unavailable
text = sample["text"] # instantiated caption
template = sample["template"] # class template
For large local scratch volumes, set the cache root before launching Python:
export STABLE_DATASETS_CACHE_DIR=~/scratch/stable-datasets
Local Data
If the Qualcomm assets are already available locally, pass data_dir=. The
directory may contain an extracted labels/ directory plus videos/, or
the original label/video archives.
ds = SomethingSomethingV2(
split="validation",
data_dir="/path/to/something-something-v2",
)
Returned Columns
video:VideoRefvideo_id: string id from the annotation filesvideo_filename: source filenamelabel: integer class id,-1when not availabletext: instantiated action texttemplate: normalized class templateplaceholders_json: JSON list of placeholder objectssplit: split name