Domain adaptive object detection via synthetically generated intermediate domain and progressive feature alignment

Abstract

The domain adaptive object detection problem is to accurately identify objects within varying target domains. The complexity arises from the discrepancies in weather conditions or diverse scenarios across different domains, which would significantly hinder the object detection model to generalize the learned knowledge from the source domain to the target domains. Currently, the teacher-student model with feature alignment is widely used to address this problem. However, most researchers only use the data from the source and target domains. To make the best use of the available data, we propose to generate the intermediate domain images by using a generative model and incorporate these images into the teacher-student model. The intermediate domain inherits the labels from the source domain and has a similar distribution to that of the target domain. To balance the influences of data from different domains on the model, we introduce the Progressive Feature Alignment (PFA) module. This strategy refines the feature alignment process. We align the source domain with the target domain by using a larger weight factor. For the intermediate domain, we use a lower weight factor for alignment with the target domain. The proposed method could significantly improve the performance of domain adaptive object detection as indicated in our experimental results: We achieve 47.9% mAP on Foggy Cityscape (from Cityscape), 63.2% AP on Cityscape (from Sim10k), and 50.6% AP on Cityscape (from KITTI).