[Paper Review] A Short Note on the Kinetics-700-2020 Human Action Dataset

Smaira, Lucas, João Carreira+4 more|arXiv (Cornell University)|Oct 21, 2020

Anomaly Detection Techniques and Applications11 references354 citations

TL;DR

The paper describes the 2020 edition of Kinetics-700, replenishing clips per class to at least 700 and provides dataset statistics, data collection details, and baseline I3D results.

ABSTRACT

We describe the 2020 edition of the DeepMind Kinetics human action dataset, which replenishes and extends the Kinetics-700 dataset. In this new version, there are at least 700 video clips from different YouTube videos for each of the 700 classes. This paper details the changes introduced for this new release of the dataset and includes a comprehensive set of statistics as well as baseline results using the I3D network.

Motivation & Objective

• Explain the motivation and updates in the 2020 Kinetics-700-2020 dataset edition.
• Present data collection improvements to address video disappearance and rare class yields.
• Provide dataset statistics, diversity analyses, and baseline I3D performance.
• Demonstrate how replenishing clips affects model training and accuracy.

Proposed method

• Describe dataset edits to ensure each class has at least 700 clips.
• Enhance text query and multilingual search to improve rare-class yield.
• Deduplicate and filter clips to remove duplicates and misclassifications.
• Report geographic distribution and diversity analyses of the final dataset.
• Evaluate baseline I3D RGB model trained from scratch on Kinetics-700-2020 with varying training sizes.

Experimental results

Research questions

• RQ1 What changes were introduced in Kinetics-700-2020 compared to prior editions?
• RQ2 How does replenishing clips per class impact dataset balance and model performance?
• RQ3 What are the yields and quality improvements for rare classes after enhanced collection methods?
• RQ4 How diverse are the videos geographically and linguistically?
• RQ5 What baseline performance does an I3D RGB model achieve on Kinetics-700-2020 across different training set sizes?

Key findings

• Kinetics-700-2020 ensures a minimum of 700 clips per class, improving balance over Kinetics-700.
• Baseline I3D RGB performance improves as the number of training examples per class increases, with top-1/top-5 scores rising across 100–600+ examples.
• I3D RGB baseline on Kinetics-700-2020 achieves 59.3%/82.0% on validation and 58.2%/80.9% on test with all training data.
• The dataset replenishment addresses video disappearance issues, maintaining high retention across splits (e.g., Kinetics-700 train 532,370 retained of 545,317; val 34,056 retained of 35,000; test 67,302 retained of 70,000).
• Multilingual and expanded text queries, plus deduplication, improve yields for rare classes (examples listed in Appendix A).
• Geographic distribution shows persistent dominance of North America, with increasing Latin American representation across editions.