Content Swapping: A New Image Synthesis for Construction Sign Detection in Autonomous Vehicles.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Autonomous vehicles require accurate detection of construction signs for safe navigation.
• Collecting sufficient real-world training data for construction sign detection is challenging due to infrequent occurrences and sign variability.
• Existing deep learning detectors struggle with limited and diverse construction sign datasets.

Purpose of the Study:

• To propose a novel method, "content swapping," for generating synthetic construction sign images.
• To enhance the training dataset for deep learning models used in autonomous driving.
• To improve the performance of construction sign detection systems for autonomous vehicles.

Main Methods:

• Content swapping divides construction signs into board (content) and frame (geometric shape) components.
• Synthetic signs are generated by combining in-domain board images with out-domain frame images.
• Generated signs are integrated into background road images using a cut-and-paste technique, with fine-tuning for realism (region, size, color).

Main Results:

• The proposed content swapping method significantly increased the number of available training images.
• The method achieved an average precision (AP50) score of 84.98% on real-world images.
• This performance surpasses off-the-shelf methods by 9.15%.

Conclusions:

• Content swapping is an effective technique for generating realistic synthetic construction signs.
• The method addresses the data scarcity problem in training deep learning models for autonomous driving.
• The released CSS138 dataset and findings benefit the autonomous driving research community.