Mutual-Assistance Learning for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection remains a challenging computer vision task.
• Current detectors struggle with non-universal features and single regression methods.
• Existing methods often yield unsatisfactory performance in complex scenarios.

Purpose of the Study:

• To propose a robust one-stage object detector named MADet.
• To address limitations of current detectors using mutual-assistance (MA) learning.
• To improve object detection accuracy, especially for objects with large aspect ratios and occlusion.

Main Methods:

• Developed MADet, a one-stage detector leveraging MA learning.
• Reintegrated decoupled classification and regression features for shared offsets in the head design.
• Employed joint anchor-based and anchor-free regression for enhanced object retrieval.
• Implemented a quality assessment mechanism for adaptive sample selection and loss reweighting.

Main Results:

• MADet achieved 42.5% AP on the MS-COCO benchmark using a ResNet50 backbone.
• The proposed method significantly outperformed existing strong baselines.
• Demonstrated state-of-the-art performance in object detection tasks.

Conclusions:

• MADet effectively addresses limitations in object detection through MA learning.
• The integrated approach enhances robustness and accuracy across diverse object characteristics.
• MADet represents a significant advancement in one-stage object detection models.