Detector With Classifier²: An End-to-End Multi-Stream Feature Aggregation Network for Fine-Grained Object Detection...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Fine-grained object detection (FGOD) combines object detection and classification.
• Existing FGOD methods often prioritize one task, limiting performance in natural scenes.
• Remote sensing images (RSIs) present unique challenges due to environmental variations.

Purpose of the Study:

• To propose a unified paradigm, Detector with Classifier2 (DC2), for end-to-end integration of object detection and fine-grained classification.
• To enhance instance localization and classification in RSIs by modeling semantic similarities.
• To effectively extract and integrate global contextual information and local intrinsic cues for improved FGOD.

Main Methods:

• Developed a two-stage sub-network: a coarse detection network and a fine-grained classification network.
• Introduced an instance-level feature enhancement (IFE) module to reduce redundant calculations and model proposal similarities.
• Proposed a multi-stream feature aggregation (MSFA) module to integrate global and local feature streams.

Main Results:

• The DC2 network achieved state-of-the-art performance on SAT-MTB and HRSC2016 datasets.
• Demonstrated significant improvements, including approximately 7% mAP gains on SAT-MTB.
• Outperformed the baseline by a substantial margin (43.2% vs. 36.7%) without complex post-processing.

Conclusions:

• The proposed DC2 paradigm effectively integrates object detection and fine-grained classification for RSIs.
• The IFE and MSFA modules contribute to enhanced feature representation and improved detection accuracy.
• DC2 offers a robust and efficient solution for fine-grained object detection in challenging remote sensing scenarios.