Dual-Resolution Dual-Path Convolutional Neural Networks for Fast Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Downsampling images speeds up visual object detection but reduces accuracy.
• Existing methods struggle to balance speed and accuracy in computationaly constrained systems.

Purpose of the Study:

• To propose DualNets, a novel dual-resolution, dual-path Convolutional Neural Network (CNN) architecture.
• To enhance accuracy in latency-sensitive visual object detection applications.

Main Methods:

• DualNets utilize two parallel CNN paths: a backbone path and an auxiliary path processing higher-resolution inputs.
• Complementary features are extracted from dual-resolution inputs.
• A progressive residual fusion strategy integrates backbone and auxiliary features for enriched representation.

Main Results:

• The DualNet architecture, integrated with the Single Shot Detector (SSD), significantly improves object detection accuracy.
• Evaluations on Pascal VOC and MS COCO benchmarks demonstrate enhanced performance.
• DualNets effectively raise accuracy in computation-payload-sensitive CNN detection applications.

Conclusions:

• DualNets offer a viable solution to the accuracy-degradation problem caused by image downsampling in object detection.
• The proposed architecture successfully balances computational efficiency with high detection accuracy.