Pixel-Wise Confidences for Stereo Disparities Using Recurrent Neural Networks

One of the inherent problems with stereo disparity estimation algorithms is the lack of reliability information for the computed disparities. As a consequence, errors from the initial disparity maps are propagated to the following processing steps such as view rendering. Nowadays, confidence measures belong to the most popular techniques because of their capability to detect disparity outliers. Recently, convolutional neural network based confidence measures achieved best results by directly processing initial disparity maps. In contrast to existing convolutional neural network based methods, we propose a novel recurrent neural network architecture to compute confidences for different stereo matching algorithms. To maintain a low complexity the confidence for a given pixel is purely computed from its associated matching costs without considering any additional neighbouring pixels. As compared to the state-of-the-art confidence prediction methods leveraging convolutional neural networks, the proposed network is simpler and smaller in terms of size (reduction of the number of trainable parameters by almost 3-4 orders of magnitude). Moreover, the experimental results on three well-known datasets as well as with two popular stereo algorithms clearly highlight that the proposed approach outperforms state-of-the-art confidence estimation techniques.