Nano-datasets: Enabling Efficient Video Understanding Research with Customizable Subsets of Large-Scale Datasets

Abstract

The advancement of self-supervised learning in video understanding has been facilitated by large-scale datasets, yet their size poses challenges for researchers with limited computational resources. To address this, we introduce nano-datasets, a repository of scripts designed to generate customizable subsets from established video datasets like Kinetics, Something-Something-v2, and ImageNet-1K. These scripts maintain the semantic integrity and structure of the original datasets while allowing users to create smaller, more manageable versions tailored to their specific research needs. By enabling researchers to experiment with diverse architectures and fine-tune models on accessible datasets, nano-datasets aims to democratize video understanding research and foster reproducibility and collaboration within the field.