Source-Free Domain Adaptation for Remote Sensing Object Detection Using Low-Confidence Pseudolabels

Abstract

Source-free domain adaptive object detection (SFOD) enables detectors trained on a source domain to be deployed to unlabeled target domains without access to the source data, thus addressing concerns about data privacy and efficiency. Existing SFOD methods typically use a mean-teacher (MT) self-training paradigm with high-confidence pseudolabels (HPLs). However, HPLs often overlook small objects in novel domain conditions, leading to biased adaptation of the student detector. This issue is particularly problematic in remote sensing (RS) datasets dominated by small vehicles. To overcome this limitation, we introduce the low-confidence pseudolabel distillation for aerial (LPLDA) scenes framework, which leverages low-confidence proposals to improve the adaptation of small objects in the target domain. Moreover, we enhance the low-confidence pseudolabel (LPL) mining process with an instance consistency (IC) loss that reinforces teacher-student consistency, making small-object features more robust to domain shifts. Extensive experiments across four practical domain shift scenarios show that our method reduces false negatives for small objects and outperforms previous SFOD approaches by effectively using domain-invariant knowledge from the source.