method: PydBox-TextDetector2019-05-28


Description: Our methods use the Fully Convolutional Network to segment text from scene image based on the EAST method. Instead of using deconvolution to enlarge the dimension of feature maps, we use a bilinear interpolation that produces softer feature maps that keep the information from the deeper layer.
By the end of the training for each image, we implement adaptive Online Hard Example Mining (OHEM) to get the hard negative and use NMS for choosing the final BBOX.