Find in Library

Water hazard detection is an important yet challenging task in autonomous driving as the complex underwater geography brings many hidden risks, e.g. puddles, which could make self-driving cars unsafe. Fully convolutional networks (FCN) have achieved remarkable performance on many image segmentation tasks, but water hazard detection problems are always hard to deal with due to the reflection characteristic of water. In this paper, we use Conditional Generative Adversarial Networks (cGAN) to deal with the water hazard detection. It has been proved that the Reflection Attention Unit (RAU) can improve the performance of deep networks for water hazard detection when added into the deep networks. We take advantage of RAU and carefully investigate its effect when placed in different layers of cGAN, with the best configuration being our proposed method: cGAN-mRAU. The `Puddle-1000' dataset is employed to evaluate our method. We use two subsets respectively and combine them together. We randomly choose some images and their ground-truth masks to train the model, and we use other images to test the model. We find many annotation mistakes in the dataset and correct them through re-annotation. Compared with FCN-8s with focal loss and 5 RAUs (FCN-8s-FL-5RAU), which is the state-of-the-art over `Puddle-1000', both cGAN and cGAN-mRAU outperform the FCN-8s-FL-5RAU in F1-measure, where cGAN achieves the best performance on `Off Road' subset and cGAN-mRAU achieves the best performance on `On Road' subset as well as whole dataset.