SSD vs. YOLO for Detection of Outdoor Urban Advertising Panels under Multiple Variabilities

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Outdoor advertisement panel detection is crucial for applications like virtual advertising and content replacement.
• Existing methods face challenges with variable panel sizes, lighting, perspectives, occlusion, and complex backgrounds.

Purpose of the Study:

• To compare the performance of Single Shot MultiBox Detector (SSD) and You Only Look Once (YOLO) deep neural networks for detecting outdoor advertisement panels.
• To evaluate these models under diverse and challenging real-world scene conditions.

Main Methods:

• Developed a custom dataset of annotated images due to the scarcity of existing data.
• Trained and evaluated both SSD and YOLO models on the custom dataset.
• Compared detection accuracy, localization precision, and false positive rates.

Main Results:

• Both SSD and YOLO achieved acceptable results across various challenging conditions.
• SSD demonstrated a significant reduction in False Positive (FP) cases.
• YOLO provided superior panel localization and a higher number of True Positive (TP) detections with greater accuracy.

Conclusions:

• YOLO is preferable for accurate outdoor advertisement panel localization and detection.
• SSD is advantageous when minimizing false positives is the primary concern.
• Further comparisons with semantic segmentation networks are included.