Stereo Confidence Estimation with Pyramid Pooling Transformer

We present a transformer-based encoder-decoder structure for estimating confidence map in stereo matching. Confidence map is obtained by estimating the reliability of the estimated disparity by stereo matching. Recently, deep convolutional neural networks (CNNs) have improved stereo estimation performance. However, these approaches have weakness with global feature. To address this limitation, we propose a novel stereo estimation method using transformer network to effectively capture the global information. Additionally, we introduce a Pyramid Pooling network to capture the local information as well as CNN structure. Experimental results show that the proposed architecture effectively capture both global and local information.