D-HAZY
Download address: http://m6z.cn/5IBatp
D-HAZY, built on the Middelbury and NYU depth datasets, which provide images of various scenes and their corresponding depth maps. A dataset of more than 1400 pairs of images, including ground truth reference images and blurred images of the same scene.
RESIDE
Download address: http://m6z.cn/5IBauH
The RESIDE dataset includes synthetic and real-world hazy images, called REalistic Single Image Dehazing, RESIDE highlights various data sources and image content, and is divided into five subsets, each of which is used for different training or evaluation purposes. A wide variety of evaluation criteria for dehazing algorithms are provided, ranging from full-reference metrics, no-reference metrics, to subjective evaluation and task-driven evaluation.
Middlebury Stereo Binocular Stereo Matching Test Dataset
Download address: http://m6z.cn/5Prq8G
These 24 datasets were created by Pan Guanghan, Sun Tiansheng, Toby Wade, and Daniel Scharstein during 2019-2021. The dataset consists of 11 scenes, imaged from 1-3 different viewing directions under many different lighting conditions and exposures (including mobile device flash and "torch" lighting).
NH-HAZE
Download address: http://m6z.cn/5tyN0D
This is a non-uniform real dataset with pairs of real haze and corresponding haze-free images. This is the first non-homogeneous image deblurring dataset, containing 55 outdoor scenes. Non-uniform fog is introduced in the scene, using a professional fog generator to simulate the real conditions of the fog scene.
DENSE-HAZE
Download address: http://m6z.cn/5tyMZP
Single image dehazing is an ill-posed problem that has recently attracted significant attention. Although interest in dehazing has increased significantly over the past few years, validation of dehazing methods remains largely unsatisfactory due to the lack of realistic haze and corresponding haze-free reference image pairs. To address this limitation, we introduce Dense Fog, a new dehazing dataset. "DENSE-HAZE" is characterized by dense and uniform hazy scenes, containing 33 pairs of real hazy images and corresponding haze-free images of various outdoor scenes. Record haze scenes by introducing real haze generated by a professional haze machine. The corresponding scenes with and without haze contain the same visual content captured under the same lighting parameters.
REVIDE Video Dehazing Dataset
Download address: http://m6z.cn/6bVqYX
The existing deep learning dehazing methods mostly use single-frame dehazing data sets for training and evaluation, so that the dehazing network can only use the information of the current foggy image to restore a clear image. On the other hand, the ideal video defogging algorithm can use adjacent foggy frames to obtain more spatio-temporal redundant information, so as to obtain a better defogging effect, but due to the lack of video defogging data sets, video Dehazing algorithms are rarely studied. To enable supervised training of video dehazing algorithms, we present for the first time a set of real video dehazing datasets (REVIDE). Using a well-designed video capture system, two captures of the same scene were successfully performed, thereby simultaneously recording paired and perfectly aligned fog and fog-free videos in the real world.