Sign in to view source links and access this dataset
Description
A subset of 9,991 video clips for few-shot learning, derived from the Kinetics-400 dataset. It contains 100 action classes and is split into 6,393 training, 1,198 validation, and 2,400 test clips. The dataset was created by StanislavLev and was last updated on July 11, 2026.
Use Cases
Benchmarking few-shot video classification models based on the provided few-shot split.
Training meta-learning algorithms for action recognition using the defined train/val/test splits.
Evaluating model generalization on a curated set of 100 human action classes.
Strengths
Contains 9,991 video clips, providing a substantial benchmark for few-shot learning.
Preserves a defined research split with 6,393 training, 1,198 validation, and 2,400 test examples.
Covers a focused set of 100 distinct human action classes.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
Derived from the Kinetics-400 dataset via the repository StanislavLev/kinetics400.
Collection Method
Built using the original metadata file few_shot_info/kinetics100.csv.
Freshness
Last updated 2026-07-11 10:26:26; freshness should be verified.
License is unknown; users should verify terms of use before application.