Learn Temporal Consistency For Robust Satellite Video Detector
Researchers have developed a new framework called Temporal Consistency Learning (TCL) designed to improve satellite video object detection. This method focuses on extracting complete, accurate, and consistent object information across entire satellite videos, particularly for oriented and fine-grained objects. TCL integrates three modules for feature aggregation, structure encoding, and temporal consistency constraints, leading to a new state-of-the-art accuracy of 47.7% mAP on the SAT-MTB benchmark, a 4.8% improvement over existing methods. AI
IMPACT Enhances object detection capabilities in satellite imagery, potentially improving surveillance and analysis applications.