A fundamental task in computer vision is the process of differentiation and identification of different objects or entities in a visual scene using semantic segmentation methods. The advancement of transformer networks has surpassed traditional convolutional neural network (CNN) architectures in terms of segmentation performance. The continuous pursuit of optimal performance, with respect to the popular evaluation metric results, has led to very large architectures that require a significant amount of computational power to operate, making them prohibitive for real-time applications, including autonomous driving. In this paper, we propose a model that leverages a visual transformer encoder with a parallel twin decoder, consisting of a visual transformer decoder and a CNN decoder with multi-resolution connections working in parallel. The two decoders are merged with the aid of two trainable CNN blocks, the fuser that combined the information from the two decoders and the scaler that scales the contribution of each decoder. The proposed model achieves state-of-the-art performance on the Cityscapes and ADE20K datasets, maintaining a low-complexity network that can be used in real-time applications.
Keywords: convolutional neural networks; deep learning; scene understanding; semantic segmentation; transformer networks.