Reducing Training Data Using Pre-Trained Foundation Models: A Case Study on Traffic Sign Segmentation Using the...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pre-trained foundation models offer adaptable solutions for specialized tasks.
• The Segment Anything Model (SAM) has shown significant capabilities in semantic segmentation.
• Data collection for model training is often resource-intensive.

Purpose of the Study:

• To evaluate if foundation models like SAM can decrease the requirement for extensive training data.
• To assess SAM's performance in traffic sign segmentation compared to established models.
• To analyze the impact of reduced training data on model accuracy.

Main Methods:

• Utilized five distinct datasets for semantic segmentation tasks.
• Focused on traffic sign segmentation for comparative analysis.
• Compared the performance of SAM against Mask R-CNN, a leading semantic segmentation model.

Main Results:

• SAM did not outperform Mask R-CNN in traffic sign segmentation, irrespective of training data volume.
• A knowledge-distilled student architecture derived from SAM demonstrated comparable accuracy.
• The SAM-derived model maintained performance even when trained on data reduced by 95%.

Conclusions:

• Foundation models like SAM may not universally surpass state-of-the-art models on all tasks.
• Knowledge distillation from SAM can create efficient models requiring significantly less training data.
• This approach holds potential for reducing data collection costs in computer vision applications.