Friday 14 March 2025
The pursuit of more accurate and efficient object detection in aerial imagery has been a long-standing challenge in the field of remote sensing. Researchers have made significant strides in recent years, but there’s still much to be desired. A new paper from Wuhan University proposes an innovative approach that leverages single-point supervision to improve object detection in aerial images.
The problem with traditional object detection methods is that they often require extensive manual annotation, which can be time-consuming and costly. Moreover, these methods typically rely on complex architectures and multiple instance learning strategies, which can lead to overfitting and decreased performance. The authors of this paper aimed to address these limitations by introducing a novel framework called PointOBB-v3.
PointOBB-v3 is designed specifically for oriented object detection in aerial images, where objects are often rotated or skewed due to the camera’s perspective. To tackle this challenge, the framework incorporates three unique image views: the original view, a resized view, and a rotated/flipped view. These views enable the model to learn more robust features that can handle varying orientations.
The key innovation in PointOBB-v3 lies in its ability to generate pseudo-rotated boxes without additional prior knowledge or annotations. This is achieved through a scale augmentation module and an angle acquisition module. The former uses a Scale-Sensitive Consistency loss function and a Scale-Sensitive Feature Fusion module to improve the model’s ability to estimate object scale, while the latter employs symmetry-based self-supervised learning to predict precise angles.
The authors evaluated PointOBB-v3 on several large-scale datasets, including DIOR-R, DOTA-v1.0/1.5/2.0, FAIR1M, STAR, and RSAR. Across all these datasets, their method achieved an average improvement in accuracy of 3.56% compared to previous state-of-the-art methods.
One notable aspect of PointOBB-v3 is its ability to adapt to different object sizes and orientations without requiring additional annotations. This makes it particularly well-suited for scenarios where objects are densely packed or have varying scales, such as in high-resolution remote sensing images.
While PointOBB-v3 demonstrates impressive performance on aerial imagery, its applicability extends beyond this domain. The framework’s reliance on single-point supervision and pseudo-rotated box generation could enable more efficient object detection in other visual domains, such as autonomous driving or surveillance.
Cite this article: “Improving Object Detection in Aerial Imagery with PointOBB-v3”, The Science Archive, 2025.
Object Detection, Aerial Imagery, Remote Sensing, Single-Point Supervision, Pseudo-Rotated Boxes, Oriented Object Detection, Scale Augmentation, Angle Acquisition, Symmetry-Based Self-Supervised Learning, Scale-Sensitive Consistency Loss Function.
