Convolutional Neural Networks or Vision Transformers: Who Will Win the Race for Action Recognitions in Visual Data?

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Action recognition in videos is a key challenge in computer vision.
• Convolutional Neural Networks (CNN) have been foundational in deep learning for visual tasks.
• Transformers, successful in NLP, are now influencing computer vision, sparking debate on their role in action recognition.

Purpose of the Study:

• To conduct a detailed study comparing CNN and Transformer models for action recognition.
• To analyze the accuracy-complexity trade-off between these two architectures.
• To discuss the future dominance of CNN or Vision Transformers (ViT) in video action recognition.

Main Methods:

• Separate analysis of CNN and Transformer models for action recognition.
• Comparative study focusing on the accuracy-complexity trade-off.
• Performance analysis to inform conclusions.

Main Results:

• Detailed comparison of CNN and Transformer performance in action recognition tasks.
• Evaluation of the balance between model accuracy and computational complexity.
• Identification of key trends and potential future directions.

Conclusions:

• The study provides insights into the comparative strengths and weaknesses of CNN and ViT for action recognition.
• The findings will guide the choice of architecture based on specific application requirements.
• Discussion on whether CNN or Vision Transformers will ultimately lead in this field.