THI Synthetic Automotive Dataset

Our THI Synthetic Automotive Dataset is built on top of CARLA and is unique since it presents varied illuminated objects (car lamps, poles, and others) and custom weather conditions, which are not present in the standard CARLA 0.9.6 version, thus presenting more challenging and diverse scenarios.

In order to ensure the generation of a diverse and challenging dataset, several criteria were established upon its creation. For the map environment, 19500 frames with resolution of 1920x1280 were collected from Towns 01 to 05, available in the native simulator which range from small scale cities, to big residential and commercial areas, as well as intercity and outer city highways. Five different light conditions were set, being morning, midday, afternoon, and night, as well as a default one which is provided by CARLA's developers for each map.

Our CARLA software was compiled from source in its 0.9.6 version and Unreal Engine 4.22 on an Ubuntu 18.04 OS. To collect the data, a tool based on Python 3.6 scripts adapted from CARLA's API was developed. This tool allows for an ease-of-use dataset creation, with the main desired features (image resolution, number of vehicles and pedestrians, and others) being integrated for user input arguments in the command line. Natively, it creates a HDF5 file with the dataset metadata, 2D bounding box of vehicles and pedestrians, and depth data, all synchronized with UNIX timestamps as indexes.

An example of the dataset can be seen in this Video.

The Dataset collection tool and respective 2D Object Detection and Monocular Depth Estimation algorithms can be accessed following the respective links.

To be able to access the dataset, please send us an email to: alessandro.zimmer@thi.de. The link to access the data will be sent to you.

If you use the Dataset, please mention publication credits.