Pyramid Geometric Consistency Learning For Semantic Segmentation

Abstract

Semantic segmentation is a critical in vision fields. Randomly transforms each image into different augmented samples and supervise the views with transformed semantics labels. However, even if the views are expanded from the same sample, the prediction results obtained by the same network will be very different. Therefore, we argue that between the augmented samples, the transformation-equivariance and the representational consistency also need to be supervised. Motivated by this, we propose a simple cross-data augmentation for semantic segmentation, in which we also leverage the pixel-level consistency constraint learning between pairs of augmented samples. As a result, our scheme significantly can improve the performances of existing semantic segmentation models without additional computation overhead. We verified the effectiveness of this method on Deeplab V3 Plus. Experiments show that our method can achieve stable performance improvement on mainstream data sets such as Pascal VOC 2012, Camvid, Cityscapes, etc.