Pothole Mix
Description
This dataset for the semantic segmentation of potholes and cracks on the road surface was assembled from 5 other datasets already publicly available, plus a very small addition of segmented images on our part. To speed up the labeling operations, we started working with depth cameras to try to automate, to some extent, this extremely time-consuming phase. We hope to be able to release a new version of the Pothole Mix dataset soon with a more significant contribution from us. In the meantime, we are releasing a set of RGB-D video pairs on which denoising/thresholding/binarization algorithms can be tested to obtain additional pairs of (image, semantic segmentation mask) to enrich the "official" dataset. The dataset reported here and as it is now (i.e. without the test set) is already able to produce fairly good models, such as the one used for the inference on this video for example: http://deeplearning.ge.imati.cnr.it/genova-5G/video/VID_20211031_162912.mp4-inference.mp4 In the readme file you can find links to all the sources that constitute the Pothole Mix dataset and the citations of the papers that published those datasets. Note: the test set has been removed from this version of the dataset to fulfill the purposes of the SHREC'22 challenge on pothole and crack detection on road pavement using RGB-D images (link: http://shrec.ge.imati.cnr.it/shrec22_road_pothole_and_crack_reco/index.html)