Misalignment-resistant domain adaptive learning for one-stage object detection

Abstract

Without consideration of task specificity, directly transforming domain adaptive pipelines from classification to one-stage detection tends to pose severer misalignments. These misalignments include: (1) Foreground misalignment that the domain discriminator obsessively concentrates on backgrounds since one-stage detectors do not contain proposals for instance-level discrimination. (2) Localization misalignment that domain-adaptive features supervised by the domain discriminator are not suitable for localization tasks, as the discriminator is a classifier in essence. To tackle these problems, we propose the Misalignment-Resistant Domain Adaption (MRDA) for one-stage detectors. Specifically, to alleviate foreground misalignment, a mask-based domain discriminator is proposed to perform instance-level discrimination by assigning the pixel-level domain labels based on instance-level masks. As for localization misalignment, a localization discriminator is introduced to learn domain-adaptive features for localization tasks. It employs an additional box-regression branch with an IoU loss to perform adversarial mutual supervision with the feature extractor. Comprehensive experiments demonstrate that our method effectively mitigates the misalignments and achieves state-of-the-art detection across multiple datasets.