The Intelligent Systems, Automation and Robotics Laboratory

Unreal Dataset

The UnrealDataset is a self-collected synthetic dataset that comprises of more than 100k images and 21 sequences collected in a bunch of highly photorealistic urban and forest scenarios with Unreal Engine and the AirSim plugin, which allows us to navigate a simulated MAV inside any Unreal scenarios. The plugin also allows us to collect MAV’s frontal camera RGB images, ground truth depth up to 40 meters and segmentation labels.


[ Training set ] [ Test set ]


In each sequence’s directory you will find: RGB images Depth ground truth: 8-bit PNG format. Here you can find the pseudocode for conversion in metric values Segmentation ground truth: two classes, “obstacle”(pixel set to 1) and “not-obstacle”(pixel set to 0) Obstacles_20m: txt files containing ground truth labels for obstacles with a depth up to 20 meters. It follows the object parametrization introduced here. The image is divided in a 8×5 grid of 32×32 cells. Each obstacle belongs to the cell containing the center of its minimum enclosing bounding box. Each line defines an obstacle and it can be read as follows:

[x-index of the cell, y-index of the cell, x coordinate of the bounding box center, y coordinates of the bounding box center, bounding box width, bounding box height, average mean of the obstacle depth, variance of the obstacle depth]

The coordinates of the bounding box center are parameterized with respect of the cell and are normalized between [0,1]. Width and height are normalized between [0,1] with respect of the size of the image. The mean and the variance are normalized between [0,1] with respect of the maximum measured depth, which is 39,75 m.

Michele Mancini