Gaussian Dynamic Convolution for Efficient Single-Image Segmentation

Interactive single-image segmentation is ubiquitous in the scientific and commercial imaging software. Lightweight neural network one practical effective way to accomplish task. This work focuses on problem only with some seeds such as scribbles. Inspired by dynamic receptive field human being’s visual system, we propose Gaussian convolution (GDC) fast efficiently aggregate contextual information for networks. The core idea randomly selecting spatial sampling area according distribution offsets. Our GDC can be easily used a module build lightweight or complex We adopt proposed address typical tasks. Furthermore, also pyramid Pooling show its potential generality common semantic segmentation. Experiments demonstrate that outperforms other existing convolutions three benchmark datasets including Pascal-Context, Pascal-VOC 2012, Cityscapes. Additional experiments are conducted illustrate produce richer more vivid features compared convolutions. In general, our conducive convolutional networks form an overall impression of image.