Data augmentation and shadow image classification for shadow detection

Shadow detection is an important branch of computer vision. Recently, convolutional neural network (CNN)-based methods for shadow have achieved better performance than based on manually designed features. However, CNNs are extremely hungry data and the training CNN-based detector requires time-consuming expensive pixel-level annotations. To alleviate this problem in detection, a method augmentation generative adversarial (GAN), named ShadowGAN, has been proposed. Given mask shadow-free image, our ShadowGAN can generate images with labels. guide get more realistic images, L 1 ${{\cal L}_1}$ loss further implemented to impose restriction between real generated images. The effectiveness demonstrated by existing detectors enlarged dataset. In addition, make use image classification task added detectors. Experiments show that feature extraction learn robust At last, these two combined achieved.