We introduce RaidaR, a new dataset that is rich in street scene images under rainy weather, and it comes with annotations in the form of both semantic and object instance segmentations; see below.
As can be seen in the figure above, the captured rainy images cover various characteristics of rainy weather conditions including fog, road reflection, water droplets, etc.
Our goal is to provide sufficient and diverse rainy scene data to augment existing datasets to improve the performance of data-driven autonomous driving algorithms. Specifically, RaidaR represents the largest such dataset to date, consisting of 58,542 rainy images, with 5,000 of them annotated with semantic segmentation and 3,658 with instance segmentation. In addition, 4,085 sunny images were also annotated with semantic segmentations.
To facilitate efficient annotation of a large volume of images in RaidaR, we develop a semi-automatic technique combining manual segmentation and an automated processing akin to cross validation. Specifically, we merge the results from four state-of-the-art segmentation networks with different weights to assess the reliability of each pixel label, where a careful manual intervention is called when there is a large inconsistency or networks predict a wrong label. On average, this approach results in a 10-20 fold speedup over completely manual segmentation.