We build a new blur detection dataset that contains 1000 images with human labeled ground-truth blur regions. These data provide useful resource to understand blur with respect to structure diversity in natural images. It enables training and testing, which is traditionally hard to implement without suitable data.
Ground truth masks are produced by 10 helpers. These images are collected from internet. They consist of images with out-of-focus blur and partial motion blur. Some sample images are shown in the above figure. The data are available for downloading below.