SOOD++: Leveraging Unlabeled Data to Boost Oriented Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Semi-supervised object detection (SSOD) commonly focuses on horizontal objects, neglecting oriented objects prevalent in aerial imagery.
• High annotation costs for oriented objects present a significant challenge in developing effective detection models.
• Existing SSOD methods do not adequately address the unique characteristics of aerial objects, such as arbitrary orientations, small scales, and dense distributions.

Purpose of the Study:

• To propose a novel Semi-supervised Oriented Object Detection (SOOD++) method to address limitations in current SSOD approaches for aerial imagery.
• To develop a technique that effectively leverages unlabeled data for improved detection of oriented objects, reducing reliance on costly annotations.
• To establish a new state-of-the-art in oriented object detection using semi-supervised learning.

Main Methods:

• A Simple Instance-aware Dense Sampling (SIDS) strategy is employed to generate comprehensive dense pseudo-labels for oriented objects.
• Geometry-aware Adaptive Weighting (GAW) loss dynamically adjusts the importance of predictions based on geometric information of aerial objects.
• Noise-driven Global Consistency (NGC) explicitly models many-to-many relationships between pseudo-labels and predictions, treating aerial images as global layouts.

Main Results:

• SOOD++ significantly outperforms previous state-of-the-art methods on DOTA-V2.0/DOTA-V1.5 benchmarks across various labeled data settings (10%, 20%, 30%).
• The method achieves substantial mAP improvements, e.g., +2.90 mAP on DOTA-V2.0 with 10% labeled data.
• SOOD++ surpasses a strong supervised baseline by +1.82 mAP, reaching a new state-of-the-art of 72.48 mAP on DOTA-V1.5.
• Demonstrates stable generalization across different oriented detectors, including multi-view oriented 3D object detectors.

Conclusions:

• The proposed SOOD++ method is effective and simple for semi-supervised oriented object detection in aerial images.
• SOOD++ successfully addresses the challenges of arbitrary orientations, small scales, and dense distributions in aerial object detection.
• The method offers significant performance gains and establishes a new state-of-the-art, highlighting the potential of semi-supervised learning in this domain.