問題：Deep convolutional neural networks have greatly improved object recognition accuracy,but remain reliant on large quantities of labeled training data.（深度卷積神經網絡依賴大量標註的訓練集）
目的：reduce annotation costs associated with detection.（減少目標檢測的標註成本）
方法：unsupervised adaptation method for object detection that combines weak global alignment with strong local alignment.
（結合弱全局對準和強局部對準的非監督自適應目標檢測模型）
具體做法：We extract global features just before the RPN and local features from lower layers, and perform weak global alignment in the high-level feature space and strong local alignment in the low-level feature space. （在RPN之前提取全局特徵，在較低層提取局部特徵，在高層特徵空間進行弱全局對準操作，在低層特徵空間進行強局部對準操作）
Weak Global Feature Alignment：
We propose to train a domain classifier to ignore easy-to-classify examples while focusing on hard-to-classify examples with respect to the classification of the domain.（我們提出訓練一個忽略容易分類的樣本，而專注於相對於域的分類難分類的樣本的域分類器）
Strong Local Feature Alignment：
The architecture of the local domain classifier, Dl, is designed to focus on the local features rather than global features. Dl is a fully-convolutional network with kernel-size equal to one. （局部域分類器專注於局部特徵而不是全局特徵，局部域分類器是尺寸爲1的全卷積網絡。）
理解：用一個分類器區分背景和目標。。。