zengxianyu/sthv2-frame-pairs

Something-Something V2 Frame Pairs Dataset

A video dataset derived from Something-Something V2 for future frame prediction tasks.

Dataset Description

This dataset contains short video clips of humans performing basic actions with everyday objects. Each video shows a simple action like "Pushing something from left to right" or "Picking something up".

Statistics

• Total videos: 168,913
• Video format: WebM
• Resolution: Variable (typically 240p)
• Duration: 2-6 seconds per clip

Dataset Structure

├── videos/
│   ├── shard_0000.tar
│   ├── shard_0001.tar
│   └── ... (47 shards total)
├── annotations/
│   └── captions.parquet
└── manifest.parquet

Files

manifest.parquet - Video metadata

Column	Description
sample_id	Unique identifier (e.g., sthv2_12345)
video_id	Original SSv2 video ID
shard	TAR shard containing this video
video_size_bytes	File size in bytes

annotations/captions.parquet - Action labels

Column	Description
sample_id	Unique identifier
caption	Action description (e.g., "Pushing something from left to right")
source	Dataset source (sthv2)

TAR Shards

Each shard contains ~3,500 videos in WebM format. Videos are named by their sample_id (e.g., sthv2_12345.webm).

Usage

import pandas as pd
from huggingface_hub import hf_hub_download
import tarfile

# Load metadata
manifest = pd.read_parquet(
    hf_hub_download("zengxianyu/sthv2-frame-pairs", "manifest.parquet", repo_type="dataset")
)
captions = pd.read_parquet(
    hf_hub_download("zengxianyu/sthv2-frame-pairs", "annotations/captions.parquet", repo_type="dataset")
)

# Merge to get full info
data = manifest.merge(captions, on="sample_id")

# Download a shard
shard_path = hf_hub_download(
    "zengxianyu/sthv2-frame-pairs",
    "videos/shard_0000.tar",
    repo_type="dataset"
)

# Extract videos
with tarfile.open(shard_path) as tar:
    tar.extractall("./videos")

Action Categories

The dataset contains 174 action categories including:

• Pushing/pulling objects
• Picking up/putting down objects
• Opening/closing containers
• Pouring liquids
• Tearing/folding materials
• And more...

License

This dataset is derived from Something-Something V2. Please refer to the original dataset terms for usage conditions.

Citation

If you use this dataset, please cite the original Something-Something V2 paper:

@inproceedings{goyal2017something,
  title={The "Something Something" Video Database for Learning and Evaluating Visual Common Sense},
  author={Goyal, Raghav and Ebrahimi Kahou, Samira and Michalski, Vincent and Materzynska, Joanna and Westphal, Susanne and Kim, Heuna and Haenel, Valentin and Fruend, Ingo and Yiber, Peter and Colombo, Manuel and others},
  booktitle={ICCV},
  year={2017}
}