Wholly-WOOD: Wholly Leveraging Diversified-Quality Labels for Weakly-Supervised Oriented Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Object detection commonly uses horizontal bounding boxes (HBoxes), but many objects require rotated bounding boxes (RBoxes) for accurate orientation estimation.
• Training oriented object detectors (OOD) typically requires costly rotation annotations.
• Existing datasets may have weaker annotations like points or HBoxes, presenting an opportunity for more efficient training.

Purpose of the Study:

• To develop a weakly-supervised oriented object detector (OOD) framework named Wholly-WOOD.
• To enable effective utilization of various annotation types, including points, HBoxes, and RBoxes, in a unified manner.
• To reduce the reliance on labor-intensive rotation annotations for training OOD models.

Main Methods:

• Developed Wholly-WOOD, a unified framework for weakly-supervised OOD.
• Implemented a method to leverage diverse labeling forms (Points, HBoxes, RBoxes) for training.
• Evaluated performance using HBox-only training against RBox-trained counterparts.

Main Results:

• Wholly-WOOD demonstrates strong performance in oriented object detection.
• Training with only HBoxes yields results comparable to RBox-trained models.
• Significantly reduces the annotation effort required for oriented object detection tasks.

Conclusions:

• Weakly-supervised training for OOD is feasible and effective.
• Wholly-WOOD offers a practical solution for leveraging existing datasets with weaker annotations.
• The framework has broad applicability in remote sensing and other domains requiring oriented object detection.