Learning contextual information for indoor semantic segmentation

Abstract

Deep Convolutional Neural Networks(DCNNs) have recently shown great performance in many high-level vision tasks, such as image classification, object detection and more recently outdoor semantic segmentation. However, the convolutional layer only process the local regions in the image, ignoring the global context information. To overcome this poor localization property of Convolutional Neural Networks(CNNs), a new form of model that combine conditional random field(CRF) to CNNs is proposed. Hence, we train the CNNs to learn local pixel-wise information and then combine the CRF based on probabilistic graph model that are connected to global pixel. The experiment results on the public indoor NYUD v2 dataset demonstrate the proposed model outperform the existing state-of-the-art methods on a challenging 40 classes task, yielding a higher class average accuracy of 47.1% and pixel average accuracy of 66.4%.