SC-RPN: A Strong Correlation Learning Framework for Region Proposal.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Current two-stage object detectors depend on region proposals for accuracy.
• Weak correlations between modules in existing region proposal methods limit performance.
• A need exists for improved region proposal techniques in object detection.

Purpose of the Study:

• To propose a novel two-stage strong correlation learning framework (SC-RPN) for enhanced region proposal generation.
• To improve the interaction and correlation among different modules within the region proposal task.
• To enhance the accuracy and effectiveness of object detection systems.

Main Methods:

• Introduced a Light-weight IoU-Mask branch for predicting intersection-over-union (IoU) masks and refining classification scores.
• Developed a Size-Aware Dynamic Sampling (SADS) strategy to ensure sampling consistency across different stages.
• Utilized point-based representation for generating region proposals with improved fitting capabilities.

Main Results:

• SC-RPN achieved a 14.5% higher Average Recall (AR₁₀₀₀) compared to the standard Region Proposal Network (RPN).
• The proposed method surpassed all existing region proposal approaches in performance.
• Integration into Fast R-CNN and Faster R-CNN yielded gains of 3.2% and 3.8% in mean Average Precision (mAP), respectively.

Conclusions:

• The SC-RPN framework effectively enhances region proposal generation through strong module correlation.
• The novel components, IoU-Mask branch and SADS, contribute significantly to performance improvements.
• SC-RPN demonstrates superior performance and effectiveness for object detection tasks.