基于单点多盒检测器的全局-局部层级的域适应目标检测

在目标检测领域里通常希望在拥有大量标记的场景中训练好的模型能够应用在无标记的其他场景中，但是不同的域分布往往是不同的，这样往往导致域迁移时模型性能的急剧下降。为了提高域迁移时模型的目标检测性能，通过两个层级来解决域迁移问题，包括全局层级迁移和局部层级迁移。这两种层级迁移分别对应不同的特征对齐方式，即全局层级采用选择性对齐方式，局部层级采用完全对齐方式。所提域迁移框架基于单点多盒检测器（SSD）模型，在全局和局部层级分别配置相应的域适配器以减少域间差异，通过对抗网络算法实现具体训练，再通过一致性正则化来进一步提高模型的域迁移性能。通过大量实验验证了提出的域迁移模型的有效性，结果表明同目前常见的域适应-快速区域卷积（DA-FRCNN）模型、对抗识别域适应（ADDA）模型以及动态对抗适应网络（DAAN）模型等三种域迁移模型相比，该模型在不同数据集上的均值平均精度（mAP）可以提高5%~10%。

Global-local domain adaptive object detection based on single shot multibox detector

In the field of object detection, it is hoped that the model trained in the domain with a lot of labels can be applied to other domains without labels, but different domain distributions are always different to each other, such difference will result in a sharp decline of model performance in domain transfer. To improve the model performance of object detection in domain transfer, the domain transfer was addressed on two levels, including the global-level transfer and the local-level transfer, which were corresponding to different feature alignment methods, that is, the global-level adopted selective alignment and the local-level adopted full alignment. The proposed domain transfer framework was constructed based on Single Shot MultiBox Detector (SSD) model and was disposed of two domain adaptors corresponding to global and local level respectively for the purpose of alleviating the domain difference. The specific training was implemented by the adversarial network algorithm, and the consistency regularization was used to further improve the domain transfer performance of the model. The effectiveness of the proposed domain transfer model was verified by many experiments. Experimental results show that on various datasets, the proposed model outperforms the existing common domain transfer models such as Domain Adaptation-Faster Region-based Convolutional Neural Network(DA-FRCNN), Adversarial Discriminative Domain Adaptation (ADDA), Dynamic Adversarial Adaptation Network (DAAN) by 5%-10% in term of mean Average Precision (mAP).