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: VideoRef

  • video_id: string id from the annotation files

  • video_filename: source filename

  • label: integer class id, -1 when not available

  • text: instantiated action text

  • template: normalized class template

  • placeholders_json: JSON list of placeholder objects

  • split: split name