Completing Missing Entities: Exploring Consistency Reasoning for Remote Sensing Object Detection.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Current remote sensing object detectors focus on rotation-invariant features, neglecting crucial spatial and semantic prior knowledge.
• This limitation hinders performance, especially with blurred or heavily occluded objects in remote sensing images (RSIs).

Purpose of the Study:

• To develop a novel framework that effectively learns and utilizes spatial and semantic prior knowledge for improved remote sensing object detection.
• To enhance the ability of object detectors to recognize objects with limited visual features, particularly in challenging conditions like occlusion and blur.

Main Methods:

• Proposed a mask-reconstruction relation learning (MRRL) framework to learn object distribution consistency through masked object reconstruction.
• Introduced a consistency-reasoning transformer over relation proposals (CTRP) to model spatial-semantic interactions and reason about difficult-to-detect objects using easier ones.
• Integrated the trained CTRP into existing detectors to boost performance on remote sensing object detection and occluded object detection tasks.

Main Results:

• Demonstrated significant improvements in object detection performance on widely-used datasets for both remote sensing and occluded object detection tasks.
• The proposed MRRL framework and CTRP component effectively leverage spatial and semantic priors, outperforming existing methods in challenging scenarios.
• The method shows robustness in detecting objects with limited visual information.

Conclusions:

• The MRRL framework and CTRP offer a powerful approach to incorporate prior knowledge into remote sensing object detection.
• The proposed method significantly enhances the detection of blurred and occluded objects, addressing a key limitation in current detectors.
• The findings suggest that modeling object relationships and employing consistency reasoning are vital for advancing object detection in complex remote sensing environments.