Explicit Visual Prompting for Universal Foreground Segmentations.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Foreground segmentation is crucial for various computer vision tasks.
• Existing methods often require domain-specific designs, limiting their generality and robustness.

Purpose of the Study:

• To introduce a unified framework for foreground segmentation tasks.
• To develop a novel visual prompting model that avoids task-specific adaptations.

Main Methods:

• Proposed Explicit Visual Prompting (EVP), inspired by NLP's pre-training and prompt tuning.
• EVP focuses tunable parameters on explicit visual content (frozen patch embeddings, high-frequency components) of individual images.
• A pre-trained model is frozen, with few extra parameters learning task-specific knowledge.

Main Results:

• EVP achieved superior performance compared to full fine-tuning and other parameter-efficient methods.
• The method demonstrated effectiveness across fourteen datasets and five distinct foreground segmentation tasks.
• Outperformed existing task-specific methods while maintaining simplicity.

Conclusions:

• EVP offers a unified, efficient, and high-performing solution for foreground segmentation.
• The framework shows scalability across different architectures, pre-trained weights, and tasks.
• This approach simplifies complex computer vision problems by leveraging explicit visual cues.