用少量真实数据提升太空目标检测模型在真实场景的精度。
Domain Adaptive Object Detection for Space Applications with Real-Time Constraints
- 结合特征不变性与域判别器，仅需250张真实标注图像即可适配。
- 在SPEED+和SPARK数据集上平均精度最高提升20点。
- 适配轻量级SSD与更先进的FCOS模型，满足实时部署需求。
目标检测在太空态势感知及相对导航等空间应用中至关重要。当前多数空间目标检测模型基于仿真数据训练，但在真实数据上性能显著下降，源于域差距。本文首次强调域适应的重要性，并探索使用极少标注真实数据的有监督域适应（SDA）方法来缩小这一差距。我们基于近期半监督适配方法，针对目标检测任务进行定制，结合域不变特征学习、基于CNN的域判别器以及域无关回归头的不变风险最小化策略。为满足实时部署要求，我们在轻量级单阶段检测器SSD（MobileNet主干）和更先进的全卷积一阶段检测器FCOS（ResNet-50主干）上测试该方法。在SPEED+和SPARK两个空间数据集上的实验表明，仅用250张标注真实图像，平均精度（AP）最高可提升20个百分点。
原文摘要 · Abstract (English)
Object detection is essential in space applications targeting Space Domain Awareness and also applications involving relative navigation scenarios. Current deep learning models for Object Detection in space applications are often trained on synthetic data from simulators, however, the model performance drops significantly on real-world data due to the domain gap. However, domain adaptive object detection is an overlooked problem in the community. In this work, we first show the importance of domain adaptation and then explore Supervised Domain Adaptation (SDA) to reduce this gap using minimal labeled real data. We build on a recent semi-supervised adaptation method and tailor it for object detection. Our approach combines domain-invariant feature learning with a CNN-based domain discriminator and invariant risk minimization using a domain-independent regression head. To meet real-time deployment needs, we test our method on a lightweight Single Shot Multibox Detector (SSD) with MobileNet backbone and on the more advanced Fully Convolutional One-Stage object detector (FCOS) with ResNet-50 backbone. We evaluated on two space datasets, SPEED+ and SPARK. The results show up to 20-point improvements in average precision (AP) with just 250 labeled real images.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。