Core Concepts
SA-MixNet proposes a novel Structure-aware Mixup and Invariance Learning framework for weakly supervised road extraction, enhancing model invariance and performance.
Abstract
Mainstream weakly supervised road extractors rely on pseudo-labels from scribbles, leading to performance degradation in varied scenes.
SA-MixNet addresses this by improving model invariance through Structure-aware Mixup and Invariance Learning.
The framework demonstrates superior performance on various datasets, outperforming state-of-the-art techniques.
Contributions include data-driven approach, improved mixup method, invariance regularization, and connectivity enhancement.
Stats
우리의 프레임워크는 다양한 데이터셋에서 우수한 성능을 보여주며 최신 기술을 능가했습니다.
SA-MixNet은 모델 불변성을 향상시키기 위해 구조적 Mixup 및 불변성 학습을 통해 이를 해결합니다.
Quotes
"Our framework demonstrates superior performance on the DeepGlobe, Wuhan, and Massachusetts datasets, outperforming the state-of-the-art techniques."
"SA-MixNet addresses the model's poor invariance to road targets in scenes with different complexities."