A weakly-supervised oriented object detector : Knowledge-based dropblock and unified regression network

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Object detection in remote sensing images (RSIs) commonly uses oriented bounding boxes (RBoxes), which are more labor-intensive than horizontal boxes (HBoxes).
• Existing weakly-supervised detectors for HBoxes often focus on discriminative object parts, reducing location accuracy.
• Spatial transformations in weakly-supervised methods create ambiguity between RBoxes and HBoxes, hindering the detection of closely situated objects.

Purpose of the Study:

• To propose a novel weakly-supervised object detector, KDUNet, that learns high-quality features and addresses the disparity between RBoxes and HBoxes.
• To enhance object localization accuracy by emphasizing the entire object rather than just discriminative parts.
• To develop a unified regression approach for both RBoxes and HBoxes to mitigate detection ambiguity.

Main Methods:

• KDUNet utilizes long-distance background information and diverse channel inputs to obscure discriminative object parts, promoting focus on the entire object.
• A novel bounding box distance measure is introduced, unifying RBoxes and HBoxes via a circumscribed rectangle and transformation angle for Gaussian distance assessment.
• The network is trained to learn high-quality feature information and compensate for the inherent ambiguity between different bounding box types.

Main Results:

• KDUNet demonstrates the capability to learn high-quality feature information and effectively reduce ambiguity in object detection.
• On the DIOR dataset, KDUNet achieved a mean Average Precision (mAP) of 57.8%, surpassing six fully-supervised networks.
• On the HRSC dataset, KDUNet achieved a mean Average Precision (mAP) of 90.1%, outperforming six fully-supervised networks.

Conclusions:

• KDUNet offers a significant advancement in weakly-supervised object detection for remote sensing images.
• The proposed methods effectively address the limitations of existing approaches, leading to improved accuracy and robustness.
• KDUNet's performance validates its potential for practical applications in remote sensing image analysis.