Properly wearing a face mask has become an effective way to limit the COVID-19 transmission. In this work, we target at detecting the fine-grained wearing state of face mask: face without mask, face with wrong mask, face with correct mask. This task has two main challenging points: 1) absence of practical datasets, and 2) small intra-class distance and large inter-class distance. For the first challenging point, we introduce a new practical dataset covering various conditions, which contains 8635 faces with different wearing status. For the second challenging point, we propose a novel detection framework about conditions of wearing face mask, named Context-Attention R-CNN, which enlarge the intra-class distance and shorten inter-class distance by extracting distinguishing features. Specifically, we first extract the multiple context feature for region proposals, and use attention module to weight these context feature from channel and spatial levels. And then, we decoupling the classification and localization branches to extract more appropriate feature for these two tasks respectively. Experiments show that the Context-Attention R-CNN achieves 84.1% mAP on our proposed dataset, outperforming Faster R-CNN by 6.8 points. Moreover, Context-Attention R-CNN still exceed some state-of-the-art single-stage detectors.

Data augmentation is an important technique to improve the performance of deep learning models in many vision tasks such as object detection. Recently, some works proposed the copy-paste method, which augments training dataset by copying foreground objects and pasting them on background images. By designing a learning-based context model to predict realistic placement regions, these approaches have been proved to be more effective than traditional data augmentation methods. However, the performance of the existing context model was limited by three problems: (1) The definitions of positive and negative samples generate too much label noise. (2) The examples with masked regions lose a lot of context information. (3) The sizes (i.e., scale and aspect ratios) of predicted regions are sampled from a prior shape distribution, which leads to a coarse estimation. In this work, we first explore the placement rules that generate realism and effective training examples for detectors. And then, we propose a trainable context model in order to find proper placement regions by classifying and refining dense prior default boxes. We also design a corresponding reasonable generation for training examples by annotating ground truth on free space according to the placement rules. The experimental results on PASCAL VOC show that our approach outperforms the state-of-the-art- related work.