Title:Fast Disparity Estimation using Dense Networks

Abstract:Disparity estimation is a difficult problem in stereo vision because the correspondence technique fails in images with textureless and repetitive regions. Recent body of work using deep convolutional neural networks (CNN) overcomes this problem with semantics. Most CNN implementations use an autoencoder method; stereo images are encoded, merged and finally decoded to predict the disparity map. In this paper, we present a CNN implementation inspired by dense networks to reduce the number of parameters. Furthermore, our approach takes into account semantic reasoning in disparity estimation. Our proposed network, called DenseMapNet, is compact, fast and can be trained end-to-end. DenseMapNet requires 290k parameters only and runs at 30Hz or faster on color stereo images in full resolution. Experimental results show that DenseMapNet accuracy is comparable with other significantly bigger CNN-based methods.