Simultaneous Deep Stereo Matching and Dehazing with Feature Attention

Abstract

Unveiling the dense correspondence under the haze layer remains a challenging task, since the scattering effects result in less distinctive image features. Contrarily, dehazing is often confused by the airlightalbedo ambiguity which cannot be resolved independently at each pixel. In this paper, we introduce a deep convolutional neural network (CNN) that simultaneously estimates a disparity and clear image from a hazy stereo image pair. Both tasks are synergistically formulated by fusing depth information from the matching cost and haze transmission. To learn the optimal fusion of depth-related features, we present a novel encoderdecoder architecture that extends the core idea of attention mechanism to the simultaneous stereo matching and dehazing. As a result, our method estimates high-quality disparity for the stereo images in scattering media, and produces appearance images with enhanced visibility. Finally, we further propose an effec

Progressive Depth Learning for Single Image Dehazing

“…Images are resized to 1280 × 384 resolution for the training and testing. Although depth estimation by stereo matching is easier than single image depth estimation, it is surprising that our method outperforms [16] in the depth estimations and dehazing tasks.…”

“…Several synthetic datasets and real-world images are utilized to compare our method against several state-of-the-art methods. The compared dehazing methods includes DCP [8], AOD-Net [10], DCPDN [20], GCANet, a very recent stereo matching and dehazing work [16]. All the reported data is tested by the authors' released code or reported results in their paper.…”

“…In outdoor applications, algorithms may fail in the extreme weather such as haze or fog. Until very recent time, Song et al [16] simultaneously performed the stereo matching and image dehazing in a feature fusion and multitask learning manner, which implicitly models the inner relationship between depth and transmission map. The stereo matching estimates the disparities from a pair of left and right cameras, and the depth is then calculated from the estimated disparities.…”

“…To train the proposed network, 6000 images are randomly selected from the training set of the KITTI raw data. In addition, 697 test images of the KITTI raw data and KITTI 2015 are used to compare our algorithm against the state-of-the-art method [16]. Images are resized to 1280 × 384 resolution for the training and testing.…”

“…It is obvious that the performance of the existed stateof-the-art depth estimation model is severely disturbed by the haze, especially for the distant areas. To firstly execute dehazing and then do depth estimation, or to fine-tune the CNNs with hazy images, is reported to be marginally helpful for stereo matching task [16]. Depth estimations for the distant area is more difficult than for the nearby regions.…”