ActionCLIP: Adapting Language-Image Pretrained Models for Video Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Natural Language Processing

Background:

• Traditional video action recognition models are limited by fixed category predictions.
• These models lack transferability to new datasets with unseen concepts.

Purpose of the Study:

• To propose a novel approach for video action recognition using multimodal learning.
• To enable zero-shot action recognition by leveraging semantic information from label texts.
• To develop a new "pre-train, adapt and fine-tune" paradigm for enhanced action recognition.

Main Methods:

• Modeling action recognition as a video-text matching problem within a multimodal framework.
• Utilizing semantic information from label texts for stronger video representation.
• Implementing a "pre-train, adapt and fine-tune" paradigm using large-scale web data.
• Introducing ActionCLIP as an instantiation of the proposed paradigm.

Main Results:

• Achieved superior and flexible zero-shot/few-shot transfer ability.
• Reached top performance on general action recognition tasks.
• Obtained 83.8% top-1 accuracy on Kinetics-400 using a ViT-B/16 backbone.

Conclusions:

• The proposed multimodal learning framework significantly enhances video action recognition.
• The "pre-train, adapt and fine-tune" paradigm offers a powerful approach for action recognition.
• ActionCLIP demonstrates strong performance and excellent transferability for zero-shot and few-shot learning.