Deep network for simultaneous stereo matching and dehazing

Unveiling the image structure and dense correspondence under the haze layer remains a challenging task, since the scattering effects cause image features to be less distinctive. In this paper, we introduce a deep network that simultaneously estimates a clear latent image and disparity from a hazy stereo image pair. To this end, inspired by a physical model of hazy image acquisition, we propose a dehazing loss function which serves as an additional cue for establishing dense correspondence. We show that stereo matching and dehazing can be synergistically formulated by incorporating depth information from haze transmission into the stereo matching process, and vice versa. As a result, our method estimates high-quality disparity for scenes in scattering media, and produces appearance images with enhanced visibility. We quantitatively evaluate the proposed method on synthetic datasets and provide an extensive ablation study. Experimental results demonstrate that our approach outperforms the recent state-of-the-art methods on both dehazing and stereo matching tasks.