YOLOH: You Only Look One Hourglass for Real-Time Object Detection

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Feature Pyramid Networks (FPN) are widely used in object detection for multi-scale accuracy.
• Traditional FPNs require extensive computation on high-resolution feature maps in the decoder.
• This computational demand limits real-time applications and efficiency.

Purpose of the Study:

• To propose a novel perspective for Feature Pyramid Networks (FPN) that reduces computational cost.
• To introduce the You Only Look One Hourglass (YOLOH) model for efficient multi-scale object detection.
• To enhance both accuracy and runtime performance compared to existing methods.

Main Methods:

• Fusing multiple feature maps into a single feature map within the encoder.
• Utilizing dense connections and dilated residual blocks to expand the receptive field of the fused feature map.
• Directly using the fused single-layer features for regression and classification.

Main Results:

• YOLOH achieves an average precision (AP) of 50.2 on the COCO dataset with a standard 3× training schedule.
• The model demonstrates a speed of 32 FPS with 40.3 AP on the ResNet-50 backbone.
• YOLOH surpasses established detector baselines in both accuracy and runtime performance.

Conclusions:

• The proposed YOLOH model offers a computationally efficient alternative to traditional FPNs for object detection.
• Early feature fusion combined with expanded receptive fields enables high performance.
• YOLOH serves as a valuable reference for designing future real-time object detection systems.