Source-free domain adaptation aims to adapt a source model trained on
fully-labeled source domain data to a target domain with unlabeled target
domain data. Source data is assumed inaccessible due to proprietary or privacy
reasons. Existing works use the source model to pseudolabel target data, but
the pseudolabels are unreliable due to data distribution shift between source
and target domain. In this work, we propose to leverage an ImageNet pre-trained
feature extractor in a new co-learning framework to improve target pseudolabel
quality for finetuning the source model. Benefits of the ImageNet feature
extractor include that it is not source-biased and it provides an alternate
view of features and classification decisions different from the source model.
Such pre-trained feature extractors are also publicly available, which allows
us to readily leverage modern network architectures that have strong
representation learning ability. After co-learning, we sharpen predictions of
non-pseudolabeled samples by entropy minimization. Evaluation on 3 benchmark
datasets show that our proposed method can outperform existing source-free
domain adaptation methods, as well as unsupervised domain adaptation methods
which assume joint access to source and target data.

本文研究基于 ImageNet 预训练特征提取器的共同学习框架，以提高面向目标域的伪标记质量并进行迁移学习，同时使用熵最小化来提高非伪标记样本的预测精度，并在三个基准数据集上进行评估。结果显示，该方法优于现有的面向源自由域适应方法和假设联合访问源和目标数据的无监督域适应方法。