The dataset was created as a part of the master thesis Deep Learning for grasssegmentation in the working space of a robotic lawn mower.
In this master thesis a deep learning approach for a robotic lawn mower will be developed.
With a binary segmentation the mower can capture the workspace own its own. As a
Data basis a whole new Dataset will be recorded from a theoretical mower perspective.
The deep learning approach will be selected from the current research, further developed
and evaluated to solve the raised problem.
If you use this dataset, please cite the thesis:
@mastersthesis{ogruhlke2018
title={Deep Learning zur Grassegmentierung im Arbeitsraum eines Rasenmähroboters},
author={Gruhlke, Oliver},
school={Universität Bremen},
year={2018}
}
The dataset consists of about 7500 images of two public parks and 16 private gardens recorded in autumn. There are different light and shadow condition as well as more and less leaves on the ground.
The appearance of the grass is a typical northern germany style. The boundaries vary over the different gardens. There are stones, fences, hedges, flowerbeds etc.
The dataset represents a typical working space of a robotic lawn mower, so the perspective is based on a possible robot. The perspective and camera parameters can be found in the masterthesis in chapter 4.
Download the dataset here.
The images are located in images folder. The files are named by the following pattern: <gardennumber>G<imagenumber>.jpg.
The ground truth for each image is located in the annotations folder. The segmentation is binary; 1 for mowable and 0 for non-mowable