Efficient and Accurate Object Detection With Asymmetric Progressive Semi-Decoupled Head and Harmonic Focal Loss

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection aims to identify objects and their locations in images.
• Current detectors struggle with simultaneous accuracy in object recognition and bounding box regression.
• Insufficient task interaction and suboptimal classification hinder detector performance.

Purpose of the Study:

• To propose a novel object detector, EADet, that improves efficiency and accuracy.
• To address the inconsistency between classification and localization in object detection.
• To enhance classifier optimization and feature representation.

Main Methods:

• Introduced Efficient Asymmetric Progressive Semi-Decoupled Head (EAPSDH) for hierarchical and dynamic task interaction.
• Developed Harmonic Focal Loss (HFL) to balance positive and negative samples and improve classifier training.
• Integrated EAPSDH and HFL into a one-stage object detector architecture.

Main Results:

• EADet effectively mitigates the inconsistency between classification and localization.
• Achieved 47.4 AP at 33.2 FPS on the MS COCO dataset using ResNet-101.
• Demonstrated a strong trade-off between accuracy and speed compared to state-of-the-art detectors.

Conclusions:

• The proposed EAPSDH and HFL significantly improve object detection performance.
• EADet offers an efficient and accurate solution for object detection tasks.
• The model shows competitive results, highlighting its practical applicability.