Uncertainty Quantification for Semi-Supervised Object Detection in Remote Sensing Images.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Semi-supervised object detection (SSOD) excels in natural scenes but is under-explored for remote sensing imagery.
• Remote sensing data presents unique challenges: arbitrary object orientations, small scales, dense distributions, and fuzzy pseudobox boundaries.
• Class imbalance and annotation difficulties hinder SSOD performance in remote sensing.

Purpose of the Study:

• To propose a novel Uncertainty Quantification (UNC) method for SSOD tailored to remote sensing images.
• To address challenges like fuzzy boundaries and class imbalance in horizontal bounding box (HBB)-based remote sensing object detection.
• To leverage unlabeled data effectively in remote sensing object detection tasks.

Main Methods:

• UNC employs uncertainty to guide network regression and classification.
• Semantic Alignment SAM Calibration (SASC) refines pseudobox boundaries using the Segment Anything Model (SAM).
• Dynamic Uncertainty Weighting (DUW) adjusts class emphasis based on instance availability and uncertainty, with a percentage threshold to prevent overemphasis.

Main Results:

• Experiments on DIOR and DOTA datasets validate UNC's effectiveness in utilizing unlabeled remote sensing data.
• UNC significantly improves detection performance compared to supervised baselines.
• On the DIOR dataset, UNC achieved mAP improvements of 12.4% (5% labeled data) and 8.6% (10% labeled data).

Conclusions:

• UNC successfully enhances SSOD for HBB-based remote sensing imagery by quantifying uncertainty.
• The method effectively tackles challenges of arbitrary orientations, small scales, dense distributions, and class imbalance.
• UNC demonstrates a robust approach to leveraging unlabeled data for improved object detection in remote sensing applications.