Self-Supervision & Meta-Learning for One-Shot Unsupervised Cross-Domain Detection

Deep detection approaches are powerful in controlled conditions, but appear brittle and fail when source models used off-the-shelf on unseen domains. Most of the existing works domain adaptation simplify setting access jointly both a large dataset sizable amount target samples. However this scenario is unrealistic many practical cases as monitoring image feeds from social media: only pretrained model available every uploaded by users belongs to different not foreseen during training. We address challenging presenting an object algorithm able exploit pre-trained perform unsupervised using one sample seen at test time. Our multi-task architecture includes self-supervised branch that we meta-train whole with single-sample cross-domain episodes, prepare condition. At deployment time task iteratively solved any incoming one-shot adapt it. introduce new media present thorough benchmark most recent methods showing advantages our approach.