The development of object detection in the era of deep learning, from R-CNN [11], Fast/Faster R-CNN [10, 31] to recent Mask R-CNN [14] and RetinaNet [24], mainly come from novel network, new framework, or loss design. However, mini-batch size, a key factor for the training of deep neural networks, has not been well studied for object detection. In this paper, we propose a Large Mini-Batch Object Detector (MegDet) to enable the training with a large minibatch size up to 256, so that we can effectively utilize at most 128 GPUs to significantly shorten the training time. Technically, we suggest a warmup learning rate policy and Cross-GPU Batch Normalization, which together allow us to successfully train a large mini-batch detector in much less time (e.g., from 33 hours to 4 hours), and achieve even better accuracy. The MegDet is the backbone of our submission (mmAP 52.5%) to COCO 2017 Challenge, where we won the 1st place of Detection task.1
Traditional steganography methods often hide secret data by establishing a mapping relationship between secret data and a cover image or directly in a noisy area, but has a low embedding capacity. Based on the thought of deep learning, in this paper, we propose a new image steganography scheme based on a U-Net structure. First, in the form of paired training, the trained deep neural network includes a hiding network and an extraction network; then, the sender uses the hiding network to embed the secret image into another full-size image without any modification and sends it to the receiver. Finally, the receiver uses the extraction network to reconstruct the secret image and original cover image correctly. The experimental results show that the proposed scheme compresses and distributes the information of the embedded secret image into all available bits in the cover image, which not only solves the obvious visual cues problem, but also increases the embedding capacity.
Understanding the change in intensity and frequency of extreme precipitation plays an important role in flood risk mitigation and water resource management in China. In this study, we analyzed the abrupt changes and long‐term trends in extreme precipitation intensity and frequency over China from 1960 to 2015 based on daily precipitation from stations. The possible teleconnection with large‐scale climate index was also been investigated. The major results are as follows: (1) 14.72% and 23.51% of all the stations over China have a change point in intensity and frequency, respectively. Moreover, most of change points occurred after 1975. (2) Extreme precipitation intensity and frequency show similar significant change trends, with a decreasing trend along the strip extending the northeast to southwest direction and an increasing trend in the two sides of the strip; 56.44% and 66.23% of all the stations shows the increasing trends in intensity and frequency. (3) Large‐scale climate indices have more influences on frequency rather than intensity. Especially, Dipole Mode Index with 1 year ahead has significantly positive correlation with extreme precipitation frequency in most areas of China, except for South China.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.