Vision-Language Models for Vision Tasks: A Survey.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional visual recognition relies on labor-intensive, task-specific deep neural network (DNN) training with crowd-labeled data.
• This paradigm is time-consuming and inefficient for the growing number of visual recognition tasks.
• Emerging Vision-Language Models (VLMs) address these limitations by leveraging large-scale image-text data.

Purpose of the Study:

• To provide a systematic review of Vision-Language Models (VLMs) for diverse visual recognition tasks.
• To analyze VLM foundations, including architectures, pre-training objectives, and downstream applications.
• To categorize and evaluate existing VLM pre-training, transfer learning, and knowledge distillation methods.

Main Methods:

• Comprehensive literature review of VLM research for visual recognition.
• Categorization of VLM pre-training strategies, transfer learning techniques, and knowledge distillation approaches.
• Benchmarking and analysis of reviewed VLM methods using widely-adopted datasets.

Main Results:

• VLMs learn rich vision-language correlations from web-scale image-text pairs.
• A single VLM can enable zero-shot predictions across various visual recognition tasks.
• This review categorizes and analyzes diverse VLM methodologies and their performance.

Conclusions:

• Vision-Language Models represent a significant advancement over traditional DNN training for visual recognition.
• VLMs offer efficient and versatile zero-shot capabilities, reducing reliance on task-specific labeling.
• Future research directions include addressing current challenges and exploring novel VLM applications in visual recognition.