TriNeXt: An efficient three-path fusion module for multi-scale feature enhancement in object detection.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Object detection frameworks require robust multi-scale feature representations.
• Enhancing spatial and semantic feature enrichment is crucial for improved detection accuracy.
• Existing methods may struggle with detecting small objects effectively.

Purpose of the Study:

• To propose a novel feature fusion module, TriNeXt, to enhance multi-scale representations in object detection.
• To integrate local, nested, and global context-aware pathways for feature enrichment.
• To evaluate the effectiveness of TriNeXt in improving object detection performance, especially for small objects.

Main Methods:

• A novel feature fusion module, TriNeXt, was designed and integrated into YOLOv5s and YOLOv8s object detection frameworks.
• TriNeXt incorporates local, nested, and global context-aware pathways.
• Experiments were conducted on the Cityscapes and KITTI datasets to evaluate performance.

Main Results:

• TriNeXt-Full improved mean Average Precision (mAP)@0.5 from 61.7% to 63.2% on YOLOv5s and from 62.3% to 63.2% on YOLOv8s.
• Recall increased from 54.5% to 55.9% (YOLOv5s) and precision rose from 77.2% to 79.2% (YOLOv5s).
• On the KITTI dataset, TriNeXt-Full improved mAP@0.5 from 94.0% to 95.0%, recall from 87.3% to 90.1%, and precision from 95.2% to 95.7%.

Conclusions:

• TriNeXt significantly enhances object detection performance, particularly for small objects, by effectively enriching multi-scale representations.
• The module achieves a favorable balance between detection accuracy and real-time inference speed.
• TriNeXt demonstrates robust generalization capabilities across diverse urban driving datasets, establishing its versatility for real-world applications.