Contrastive Learning for Improving the Robustness in Temporal Action Localization
Weijuan Li · 2023
This paper presents introducing contrastive learning to the model for temporal action localization (TAL) tasks, inspired by its extremely wide applications, including natural language processing (NLP), computer vision (CV), automatic speech recognition (ASR) and so on. While the applications in CV domain mainly focus on image representations learning tasks such as object recognition and image retrieval, we propose to apply the contrastive learning to TAL. We have primarily done two parts, data augmentation and contrastive loss, which are both crucial to contrastive learning. For data augmentation,we adopt cropping and color distortion to generate positive and negative pairs which are essential components for training of contrastive learning model. In addition, we build a formula Lc representing the contrastive loss, through which the model can learn how to map similar samples to nearly the same embedding space while mapping dissimilar samples to distant embedding spaces. What we do contributes to improving the accuracy and generalized ability of the model.