The ForestSim dataset focuses on data collection from unstructured simulated evironments. This is to help improve semantic understanding of unstructured outdoor environments for applications of off-road autonomous navigation. The data is collected from 25 different simulated environments built using the Unreal Engine. These environments meet the criteria for unstructured enviornments which are classified as rough unstructured terrain with large variations in geometry, appearance, and shapes where tall grass and rough terrain may be interpreted as non-navigable terrain. In unstructured areas there is more ambiguity in visual perception.
Some of the environments used here are semi-structured but these were very large environments where there were areas that fit our criteria. These environments were selected based upon a predetermined criteria where the environment contains large variations in geometry, appearance, and shapes where tall grass and rough terrain.
These images were collected using Airsim and Unreal Engine. The data required a post processing pipeline to create ground truth labels. The techniques applied were to first determine the object classes to focus on for the dataset and then assign them a rgb value. Data collected from the various environment were manually curated to convert the pixels of the corresponding object to the pre-determined rgb value. The final ground truth segmentation images are provided here but please note they required processing and cleaning after collection with Airsim from Unreal Engine.
The github repository associated with ForestSim is located at ForestSim Github
Downloads are available for the collected raw rgb data, the post processed rgb segmentation data used for training and testing, and the mapping used of rgb values to object
The sample contains 21 raw rgb images and 21 rgb segmentation images
The total data consists of 2094 rgb raw images and 2094 rgb segmentation images
When using this dataset in your research, we would appreciate if you would cite us! [paper]
@inproceedings{ForestSim,
author = {Wagle, Pragat and Chen, Zheng and Liu, Lintao},
title = {ForestSim: A Synthetic Benchmark for Intelligent Vehicle Perception in Unstructured Forest Environments},
booktitle={IEEE Intelligent Vehicles Symposium (IV)},
year = {2026}
}