Core Concepts
The author proposes a novel active learning method, SSOD-AT, to enhance semi-supervised object detection in remote sensing images using a teacher-student network. By incorporating a RoI Comparison module and diversity criterion, the method outperforms state-of-the-art techniques.
Abstract
The lack of object-level annotations in remote sensing images poses challenges for object detection. The proposed SSOD-AT method combines active learning and semi-supervised learning to improve annotation quality. It introduces a RoI Comparison Module (RoICM) to generate high-confidence pseudo-labels and selects top-K uncertain images for human labeling. The method integrates uncertainty and diversity criteria to enhance the active learning process. Experimental results on DOTA and DIOR datasets show superior performance compared to existing methods.
Stats
Compared with the best performance in the SOTA methods, the proposed method achieves 1% improvement at most cases in the whole active learning.
DOTA dataset contains 2806 aerial images from different sensors with 15 common object categories.
DIOR dataset consists of 23,463 images manually labeled with axis-aligned bounding boxes covering 20 object categories.
Quotes
"The proposed method can provide both confident pseudo-labels and informative images."
"A RoI comparison module(RoICM) is introduced by comparing the RoIs generated by teacher and student network."
"The combination of the two sampling strategies maximizes the effectiveness of AL process."