A Multipath Fusion Strategy Based Single Shot Detector.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection is crucial for intelligent systems and sensor applications.
• One-stage detectors offer efficiency for real-time processing compared to two-stage methods.
• Existing one-stage detectors require accuracy improvements for complex tasks.

Purpose of the Study:

• To enhance the accuracy of one-stage Single Shot Detectors (SSD).
• To introduce a novel Multi-Path fusion Single Shot Detector (MPSSD).
• To improve object detection performance with minimal computational overhead.

Main Methods:

• Proposed a novel Multi-Path fusion Single Shot Detector (MPSSD).
• Exploited pyramid-like connections among different scale representations.
• Introduced feature fusion and pyramid aggregation modules for enhanced feature generation.

Main Results:

• MPSSD demonstrated superior performance over state-of-the-art detectors on benchmark datasets (PASCAL VOC2007, VOC2012, MS COCO).
• Achieved high mean Average Precision (mAP): 81.8% on VOC2007, 80.3% on VOC2012, and 33.1% on COCO test-dev 2015 for 512x512 images.
• Enhanced features incorporated both localization and semantic information, improving detection accuracy.

Conclusions:

• The proposed MPSSD effectively improves object detection accuracy and efficiency.
• The novel feature fusion strategy offers a significant advancement in one-stage detection.
• MPSSD provides a computationally efficient solution for real-time object detection applications.