A Survey of Visual Transformers.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Transformer architectures, initially successful in Natural Language Processing (NLP), are increasingly adapted for Computer Vision (CV).
• Visual Transformers (ViTs) leverage attention mechanisms for image analysis, demonstrating competitive performance against Convolutional Neural Networks (CNNs).

Purpose of the Study:

• To provide a comprehensive survey of over 100 visual Transformer models.
• To organize and analyze existing ViTs based on CV tasks and data modalities.
• To identify unexploited research avenues and suggest future directions for visual Transformers.

Main Methods:

• Systematic review and taxonomy of visual Transformer architectures.
• Categorization based on fundamental CV tasks (classification, detection, segmentation) and data streams (images, point clouds, vision-language).
• Comparative evaluation of ViTs under various configurations and benchmarks.

Main Results:

• Visual Transformers achieve state-of-the-art performance across multiple CV benchmarks.
• A proposed taxonomy organizes methods by motivation, structure, and application.
• Identified potential improvements, such as semantic embeddings, to enhance ViT capabilities.

Conclusions:

• Visual Transformers represent a significant advancement in computer vision.
• Further research into areas like semantic embeddings can unlock greater potential.
• The survey provides a foundation for future development and investment in ViT research.