Road Segmentation

We observe that satellite imagery is a powerful source of information as it contains more structured and uniform data, compared to traditional images. Although computer vision community has been accomplishing hard tasks on everyday image datasets using deep learning, satellite images are only recently gaining attention for maps and population analysis. This workshop aims at bringing together a diverse set of researchers to advance the state-of-the-art in satellite image analysis. To direct more attention to such approaches, we propose DeepGlobe Satellite Image Understanding Challenge, structured around three different satellite image understanding tasks. The datasets created and released for this competition may serve as reference benchmarks for future research in satellite image analysis. Furthermore, since the challenge tasks will involve "in the wild" forms of classic computer vision problems, these datasets have the potential to become valuable testbeds for the design of robust vision algorithms, beyond the area of remote sensing.

The datasets introduced in Chapter 6 of my PhD thesis are below. See the thesis for more details. If you use any of these datasets for research purposes you should use the following citation in any resulting publications: @phdthesis{MnihThesis, author = {Volodymyr Mnih}, title = {Machine Learning for Aerial Image Labeling}, school = {University of Toronto}, year = {2013} }

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. The main dataset is composed of 4340 (image,mask) pairs at different resolutions divided into training/validation/test sets with a proportion of 3340/496/504 images equal to 77/11/12 percent. This is the dataset used in the SHREC2022 competition and it is the dataset that allowed us to train the neural networks for semantic segmentation capable of obtaining the nice images and videos that you have probably already seen. Along the main dataset we also release a set of RGB-D videos consisting of 797 RGB clips and as many clips with their disparity maps, captured with the excellent OAK-D cameras we won for being finalists at the OpenCV AI Competition 2021. In an effort to achieve (semi-)automatic labeling for these clips, we postprocessed the disparity maps using classic CV algorithms and managed to obtain 359 binary mask clips. Obviously these masks are not perfect (they cannot be by definition, otherwise the problem of automatic road damage detection would not exist), but nonetheless we believe they are an excellent starting point to create, for example, new data augmentations (creating potholes on "intact road images" belonging to other standard road datasets) or to be used as textures in the creation of 3D scenes from which to extract large amounts of images/masks for the training of neural networks. You can have a preview of what you will find in these clips by watching this video showing the overlay of images and binary masks: http://deeplearning.ge.imati.cnr.it/genova-5G/video/pothole-mix-videos/pothole-mix-rgb-d-overlay-videos-concat.html Please take a look at the readme file inside the main dataset zipfile to have some more details about the single sub-datasets and their sources.

A novel remote sensing dataset for evaluating a geospatial machine learning model's ability to learn long range dependencies and spatial context understanding. We create a task to use as a proxy for this by training models to extract roads which have been broken into disjoint pieces due to tree canopy occluding large portions of the road. The dataset consists of 30,000 RGBN NAIP images and land cover annotations from the Chesapeake Conservacy containing significant amounts of the Tree Canopy Over Road category.

The dataset includes polarimetric, RGB and depth automotive (on the road) data. The data includes synchronized and aligned samples of the following: angle of linear polarization (AoLP) images, degree of linear polarization (DoLP) images, RGB images, lidar scans, ground truth free space segmentation (road segmentation), GNSS / IMU readings (vehicle location, vehicle orientation, vehicle speed, vehicle acceleration, etc.) and calibration matrices. The dataset includes 12,627 samples from 6 different locations. The locations represent typical suburbs where the scenes are not remarkably cluttered, but provide a good distribution of vehicles, pedestrians and buildings. Additionally, the dataset includes free space segmentation of 8,141 images.