Domain-adaptive Instance Segmentation for Visual Inspection from Civil Infrastructure Images

Abstract

This paper focuses on optimizing domain adaptation for instance segmentation to address the rapid adaptation of visual inspection on an unseen target domain. We propose enhancing Mask R-CNN with a domain adaptation capability. The proposed model performs domain adaptation by controlling the prediction to be target domain-specific with a latent vector, representing the domain property. Since the latent vector can be inferred from sets of unlabeled images in a domain by Deep sets, the proposed model provides a domain adaptation capability without the need to either retrain the model on the target domain or prepare labeled target data.