Multi-factor consideration sample assignment for oriented tiny object detection.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Tiny object detection in aerial imagery is vital for urban planning and environmental monitoring.
• Challenges include unpredictable object orientation and lack of distinctive features, leading to sample assignment mismatches.
• Existing methods struggle with accurate anchor-prior association for oriented objects.

Purpose of the Study:

• To introduce a novel sample assignment mechanism for improved tiny object detection in aerial images.
• To address challenges in assigning positive samples for objects with varying orientations.
• To enhance the precision of aerial object detection systems.

Main Methods:

• Developed a multi-factor consideration sample assignment (MCSA) mechanism.
• Introduced a dynamic prior block (DPB) for dynamic alignment of priors with objects.
• Implemented an anchor assessment metric and a dynamic Gaussian mixture model (DGMM) for refined sample selection.

Main Results:

• The MCSA method demonstrated superior performance over existing techniques on DIOR-R, DOTA-v2.0, DOTA-v1.5, and DOTA-1.0 datasets.
• Achieved a mean Average Precision (mAP) of 51.86% on the DOTA-v2.0 dataset, a 5.18 percentage point improvement over the baseline.
• The approach ensures precise sample assignment through dynamic prior distribution and selection of compatible positive samples.

Conclusions:

• The proposed MCSA mechanism effectively improves tiny object detection accuracy in aerial imagery.
• Dynamic prior alignment and robust sample selection are key to enhancing detection precision for oriented objects.
• This work offers a significant advancement in aerial image analysis for applications like urban planning and environmental monitoring.