Visual cross-domain adaptation under various data access privileges
Deep neural networks have achieved promising performance on solving multiple computer vision tasks when trained on large-scale in-domain datasets. For some specific learning tasks, the model learned from the off-the-shelf well-labeled training set suffers from performance degradation when evaluated on the novel test set due to their distribution divergence. The practical dilemma motivates the emerging research topic named as ``Unsupervised Domain Adaptation''. This dissertation is centered with a novel perspective of cross-domain data access privileges to discuss various domain adaptation scenarios including unlimited cross-domain data access, source-data absent scenario and target-data missing scenario. The first scenario is unsupervised domain adaptation where training model utilizes well-labeled source domain and unlabeled target samples. This condition is beneficial to learn domain-invariant representation with all available data. To mitigate domain shift, we propose a structural preserving generative method to perform graph alignment. Moreover, this thesis also considers a novel metric with cross-domain graph information to implement category-wise alignment. The second scenario is source-free domain adaptation, where only a well-trained source model and target data are available for knowledge transfer and adaptation. To address this, we propose an adaptive adversarial network to improve classification ability and transfer source knowledge by developing a flexible target-specific classifier. To promote the model's robustness, the approach utilizes category-wise matching and self-supervised learning. Finally, the target-data missing scenario comprises two sub-problems. The first sub-problem is when there are no target instances for learning the model. In this case, domain generalization (DG) can be used, which introduces multiple source domains to learn domain-invariant features. Furthermore, we extend DG by exploring the change of model generalization ability with imbalanced data distribution and use data augmentation to overcome it. The other strategy is zero-shot domain adaptation, which utilizes an additional task-irrelevant dataset to learn cross-domain contents and improve model generalization ability. Additionally, we formulate the second sub-problem as incomplete multi-view domain adaptation, where the multi-view source data and single-view target instances are available for model training. We adopt channel-wise change and enhancement to recover missing information and align various distributions.