Integrated Multiscale Domain Adaptive YOLO.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Domain shift is a key challenge in deep learning, where training data distributions differ from real-world testing data.
• Existing object detection models often struggle with performance degradation due to domain shift.

Purpose of the Study:

• To introduce a novel MultiScale Domain Adaptive YOLO (MS-DAYOLO) framework to address domain shift in object detection.
• To propose three new deep learning architectures for Domain Adaptation Networks (DANs) to generate domain-invariant features.

Main Methods:

• Developed MS-DAYOLO framework integrating multiple domain adaptation paths and classifiers with YOLOv4.
• Proposed three novel DAN architectures: Progressive Feature Reduction (PFR), Unified Classifier (UC), and Integrated.
• Trained and tested DAN architectures with YOLOv4 on popular datasets for autonomous driving applications.

Main Results:

• Significant improvements in object detection performance for YOLOv4 when tested on target data.
• MS-DAYOLO achieved comparable object detection performance to Faster R-CNN.
• MS-DAYOLO demonstrated an order of magnitude real-time speed improvement over Faster R-CNN.

Conclusions:

• The proposed MS-DAYOLO framework effectively mitigates domain shift issues in object detection.
• MS-DAYOLO offers a real-time, high-performance solution for autonomous driving applications.
• The novel DAN architectures contribute to generating robust, domain-invariant features for improved detection.