Metric-Based Few-Shot Learning for Video Action Recognition

TL;DR

This paper explores metric-based few-shot learning methods for video action recognition, demonstrating that two-stream models with prototypical networks and LSTM encoders achieve high accuracy on Kinetics 600 with limited labeled examples.

Contribution

It introduces a comprehensive evaluation of metric-based few-shot learning approaches for video classification, highlighting the effectiveness of two-stream models and specific encoder architectures.

Findings

Two-stream models significantly improve performance.

Prototypical networks outperform other few-shot algorithms.

Achieved 84.2% accuracy on standard 5-shot 5-way task.

Abstract

In the few-shot scenario, a learner must effectively generalize to unseen classes given a small support set of labeled examples. While a relatively large amount of research has gone into few-shot learning for image classification, little work has been done on few-shot video classification. In this work, we address the task of few-shot video action recognition with a set of two-stream models. We evaluate the performance of a set of convolutional and recurrent neural network video encoder architectures used in conjunction with three popular metric-based few-shot algorithms. We train and evaluate using a few-shot split of the Kinetics 600 dataset. Our experiments confirm the importance of the two-stream setup, and find prototypical networks and pooled long short-term memory network embeddings to give the best performance as few-shot method and video encoder, respectively. For a 5-shot 5-way task, this setup obtains 84.2% accuracy on the test set and 59.4% on a special "challenge" test set, composed of highly confusable classes.