Instance-Level Orientation Enhancement for Horizontal Box Supervised Oriented Object Detection in Remote Sensing...

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Remote sensing datasets predominantly use horizontal bounding boxes (HBBs), creating a mismatch with oriented object detection methods requiring oriented bounding boxes (OBBs).
• Existing horizontal box supervised oriented object detection methods face limitations in learning instance-specific orientation features due to image-level transformations and ineffective feature extraction.
• Conventional methods often rely on data augmentation for orientation awareness but fail to extract orientation-sensitive features effectively using standard CNNs.

Purpose of the Study:

• To address the limitations of existing methods in horizontal box supervised oriented object detection.
• To propose a novel detector, the Instance-Level Orientation Information Enhanced Detector (ILOEDet), to improve orientation sensitivity.
• To enhance the model's ability to learn instance-specific orientation features and decouple object orientation from global image context.

Main Methods:

• Integration of the Instance-Aware Rotated Convolution Module (IARCM) for embedding instance-level orientation information.
• Implementation of an Instance-Level Flip Consistency (IFC) mechanism for robust orientation learning through self-supervised learning.
• Leveraging classification and center-ness scores within IARCM to select high-quality instances and guide rotated convolution operations.

Main Results:

• The proposed ILOEDet effectively improves orientation sensitivity in object detection.
• Experiments on DOTA, HRSC2016, and DIOR-R datasets demonstrate the superior performance of the proposed approach.
• The Instance-Aware Rotated Convolution Module and Instance-Level Flip Consistency mechanism contribute to enhanced feature extraction and orientation learning.

Conclusions:

• ILOEDet offers a significant advancement in horizontal box supervised oriented object detection.
• The proposed instance-level learning mechanisms overcome the limitations of image-level transformations and conventional CNNs.
• The method provides a more robust and effective solution for oriented object detection in remote sensing applications.