HTD: Heterogeneous Task Decoupling for Two-Stage Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Two-stage object detectors often suffer from task misalignment between classification and regression heads.
• Existing decoupling methods use similar structures, neglecting task-specific feature needs and shared semantic knowledge.

Purpose of the Study:

• To propose a Heterogeneous Task Decoupling (HTD) framework that effectively addresses task misalignment and semantic inconsistency in object detection.
• To introduce novel modules for specialized feature extraction and knowledge sharing between decoupled task heads.

Main Methods:

• Developed a Semantic Feature Aggregation (SFA) module to capture global semantics as shared knowledge.
• Introduced a Progressive Graph (PGraph) module for enhanced region proposal representation through graph reasoning for classification.
• Designed a Border-aware Adaptation (BA) module to integrate multi-level features for regression, focusing on border perception.

Main Results:

• HTD framework significantly outperforms existing object detection methods.
• Achieved state-of-the-art single-model performance with 50.4% AP and 33.2% APs on the COCO test-dev set using a ResNet-101-DCN backbone.
• Demonstrated improved instance-level semantic consistency (ISC) through knowledge aggregation.

Conclusions:

• The proposed HTD framework effectively decouples heterogeneous tasks in object detection, leading to superior performance.
• Specialized modules (PGraph, BA) and shared knowledge aggregation (SFA) are crucial for overcoming limitations of previous decoupling strategies.
• HTD sets a new benchmark for object detection accuracy and efficiency.