Something-Something V2

• Modality: RGB video (crowdsourced, 2-6 seconds)
• Primary Tasks: Fine-grained action recognition, temporal reasoning, few-shot transfer
• Scale: 220,847 video clips, 174 action templates, 338,000 textual descriptions
• License: Apache License 2.0
• Access: https://developer.qualcomm.com/software/ai-datasets/something-something

Summary

Something-Something V2 focuses on human-object interactions captured with short, templated instructions. Its fine-grained verb-noun combinations drive research in temporal reasoning, compositional action recognition, and grounding natural language descriptions to video.

Reference Paper

• R. Goyal et al. "The 'Something Something' Video Database for Learning and Evaluating Visual Common Sense." ICCV, 2017. PDF

Benchmarks & Baselines

• Temporal Shift Module (TSM) - Top-1: 63.4, Top-5: 89.5; Lin et al., ICCV 2019.
• ViViT-L/16x2 - Top-1: 68.8, Top-5: 90.6; Arnab et al., ICCV 2021.
• Standard split includes train, validation, and test; test labels withheld for leaderboard submission via EvalAI.

Tooling & Ecosystem

• EvalAI leaderboard for official metrics.
• PyTorchVideo and MMAction2 supply dataloaders.
• something-something-downloader CLI to fetch videos and metadata.

Known Challenges

• Video downloads occasionally fail due to host availability; rerun downloader scripts.
• Labels are heavily templated-consider text augmentation to avoid overfitting to phrasing.
• Clips are low-resolution (240p); models benefit from temporal cues rather than spatial detail.

Cite

@inproceedings{goyal2017something,
  title     = {The `Something Something' Video Database for Learning and Evaluating Visual Common Sense},
  author    = {Goyal, Raghav and Kahou, Samira Ebrahimi and Michalski, Vincent and others},
  booktitle = {Proceedings of the IEEE International Conference on Computer Vision},
  year      = {2017}
}