An overview of the GOS dataset, which is captured based on the roads of GTA V,
contains various scenes and plentiful attributes.
|
GOS: GTA V Outdoor Scene Dataset |
|
|
|
An overview of the GOS dataset, which is captured based on the roads of GTA V, |
 Paper |
 Statistics |
 Code |
Grand Theft Auto V (GTA V) is an action-adventure video game published by Rockstar Games. GTA V applies the proprietary engine to improve rendering qualities, makes its scene rendering, texture details, light and shadow effects are comparable to the realworld scenes.
We develop a tool for outdoor scenes collecting in GTA V based on Script Hook V. In order to acquire depth and stencil map corresponding to scene images, we utilize tool provided by GTAVisionExport.
We generate the GTA V Outdoor Scene dataset (GOS) using the built tool, which provides rich variety scene images and coarse-to-fine attribute annotations, including viewpoint, distance, scene, weather and illumination. The details will be introduced in the following section.
GOS dataset contains 5 different types of viewpoints, every 30° from -60° to 60° on each side of roads.
And 5 different types of distance: 0m, 10m, 15m, 25m, 50m.
|
GOS dataset owns various backgrounds, including urban areas and wild scenes, such as desert, forest, beach, highway and mountain, etc. Bold and underline represents the properties included in the dataset.
|
We present GOS dataset consisting of 8 different types of time, e.g. midnight (0:00), predawn (5:00), dawn (6:00), morning (8:00), midday (12:00), afternoon (16:00), sunset (18:30) and dusk (21:00).
|
We present GOS dataset consisting of 8 different types of weather, e.g. sunny, clear, cloudy, foggy, overcast, rainy, stormy, snowy.
|
GOS dataset consists of 4,632,500 images, contains 8149 places and 231,625 scenes.
The original resolution of dataset is 1920*1080. Below is a set of sample images, contains color image, depth map and stencil map.
Color image |
Depth map |
Stencil map |
In order to match the GAN training, we also privide another version which crop images along the center position as a square, and scales to 256*256. The most images displayed in the website is suitable for this format.
Viewpoint |
Distance |
We present a detailed comparison between our GOS dataset and existing datasets,
including data volume, image resolution, annotation types.
GTA V contains 226 main roads distributed in cities and rural areas.
We have obtained several the information about the road in GTA V, including start and end points, waypoints, and road types.
The collection points are set up every 15 meters on each road.
 |
We arrange multiple cameras with different positions at each collection point, which contains five different distances
(0m, 10m, 15m, 25m, 50m) and five different viewpoints (-60°, -30°, 0°, 30°, 60°).
Cameras are in the same horizontal plane with the collection point.
 |
When the camera gradually moves away from the collection point, it may be occluded by objects beside the street, or even enters a building or mountain. Therefore, it is necessary to exclude unreasonable scenes.
Depth map and stencil map is utilized to filter unqualified images. The vehicle in the image is placed on the collection point as a baseline for filtering. The following are a few typical situations.
Scene occlusion |
Render missing |
@inproceedings{xie2022gos,
title={GOS: A Large-Scale Annotated Outdoor Scene Synthetic Dataset},
author={Xie, Mingye and Liu, Ting and Fu, Yuzhuo},
booktitle={ICASSP 2022-2022 IEEE International Conference on Acoustics
, Speech and Signal Processing (ICASSP)},
pages={3244--3248},
year={2022},
organization={IEEE}
}
Rockstar Games allows the players to develop the mod for noncommercial or personal use, and can only be used in offline version.
We will release the GOS dataset after all work is completed.
The template of this webpage is borrowed from Suncheng Xiang.
For further questions and suggestions, please contact Mingye Xie (xiemingye@sjtu.edu.cn).