Supervised Contrastive Learning for Few-Shot Action Classification

Abstract

In a typical few-shot action classification scenario, a learner needs to recognize unseen video classes with only few labeled videos. It is critical to learn effective representations of video samples and distinguish their difference when they are sampled from different action classes. In this work, we propose a novel supervised contrastive learning framework for few-shot video action classification based on spatial-temporal augmentations over video samples. Specifically, for each meta-training episode, we first obtain multiple spatial-temporal augmentations for each video sample, and then define the contrastive loss over the augmented support samples by extracting positive and negative sample pairs according to their class labels. This supervised contrastive loss is further combined with the few-shot classification loss defined over a similarity score regression network for end-to-end episodic meta-training. Due to its high flexibility, the proposed framework can deploy the latest contrastive learning approaches for few-shot video action classification. The extensive experiments on several action classification benchmarks show that the proposed supervised contrastive learning framework achieves state-of-the-art performance.