Convolutional Networks with Bracket-Style Decoder for Semantic Scene Segmentation

To build up a state-of-the-art semantic scene segmentation model, a balanced combination between coarsely and finely contextual details is required for eliminating class-wise ambiguities and reaching high accuracy of pixel-wise labeling, respectively. Accordingly, with deep learning integration, prior works have achieved impressive performance in general, but found difficulties in correctly labeling medium to small objects. For the purpose of overcoming such issue, this paper proposes a deep convolutional network with bracket-style decoder, namely B-Net, to leverage the utilization of features learned at middle layers in the backbone networks (encoder) for constructing a final prediction map of densely enhanced semantic information. In particular, every feature map of interest combines with its adjacent version of higher spatial resolution through lateral connection modules to produce finer outputs that repeat such routine round-by-round until retrieving the finest-resolution map for dense prediction. Consequently, benchmarking results on CamVid dataset showed the effectiveness of the proposed method with mean class-wise accuracy, pixel-wise accuracy, and mean union intersection of 76.2%, 87.1%, and 66.4%, respectively.