Padding Investigations for CNNs in Scene Parsing Tasks

Zero padding is widely used in convolutional neural networks (CNNs) to prevent the size of feature maps diminishing too fast. However, it has been claimed to disturb the statistics at the border [9]. In this work, we compare various padding methods for the scene parsing task and propose an alternative padding method (CApadding) by extending the image to alleviate the border issue. Experiments on Cityspaces [2] and Deep-Globe [3] show that models with the proposed padding method achieves higher mean Intersection-Over-Union (IoU) than the zero padding based models.