Interactive Regression and Classification for Dense Object Detector.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection performance relies heavily on feature representation.
• Localization information, crucial for detection, is often underutilized in current models.
• Existing methods struggle to effectively integrate regression and classification features.

Purpose of the Study:

• To propose a novel method, Interactive Regression and Classification (IRC), for improved object detection.
• To enhance the utilization of localization information within object detection frameworks.
• To develop an efficient and integrable approach for boosting detection accuracy.

Main Methods:

• Introduced the Feature Aggregation Module (FAM) and Localization Attention Module (LAM) to leverage localization information for classification.
• Implemented an interactive learning process where the classifier guides the regression branch during backward propagation.
• Designed IRC for seamless integration into both anchor-based and anchor-free object detectors.

Main Results:

• Achieved 47.2% AP on COCO test-dev using a ResNet-101 backbone, surpassing state-of-the-art methods like PAA and GFL.
• Demonstrated significant performance improvements across various popular dense object detectors.
• Attained a single-model single-scale AP of 51.4% with a Res2Net-101-DCN backbone.

Conclusions:

• IRC effectively enhances object detection by interactively learning regression and classification features.
• The proposed method offers substantial performance gains without compromising computational efficiency.
• IRC presents a versatile and effective solution for advancing object detection technology.