What is TartanAir?
Overview of TartanAir
TartanAir is a dataset designed for robot navigation tasks. It is developed by researchers at the Robotics Institute of Carnegie Mellon University. Data for TartanAir is collected in photo-realistic simulation environments. The dataset was presented at the 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). The TartanAir dataset is at least 1 order of magnitude larger than existing datasets. The total size of the TartanAir dataset is up to 3TB. The dataset is hosted on servers in the United States.
Data Composition and Features
The TartanAir dataset contains over 1 million frames of data for visual SLAM research. The dataset includes 1037 long motion sequences. The dataset includes multi-modal sensor data including RGB stereo images, depth images, segmentation labels, and camera poses. TartanAir includes ground truth labels such as optical flow, LiDAR, and stereo disparity. TartanAir camera poses are defined in the NED frame. The dataset utilizes the Unreal Engine for realistic 3D scene rendering. TartanAir scenes feature varied conditions such as shadows, over-exposure, and reflections of glass or puddles. The dataset includes data collected at different times of day and in different seasons. TartanAir includes light sources ranging from road lamps and neon lights to sunlight and moonlight. The dataset includes scenarios with adverse weather such as rain, snow, and fog. The dataset contains dynamic objects such as simulated humans, vehicles, and machinery. In the experiment, dynamic objects were configured to different levels of dynamics. The dataset provides data categorized by motion pattern difficulty, specifically Easy and Hard. TartanAir's motion pattern is described as more diverse compared to the KITTI dataset.
Technical Implementation and Usage
The dataset is used for the CVPR Visual SLAM challenge. The TartanAir dataset was used to train the generalizable learning-based visual odometry system called TartanVO. Evaluation for TartanAir adopts ATE and RPE metrics. The mapping process for an environment size of 100m x 100m x 10m takes about 1 hour. The RRT* planning algorithm is used in TartanAir to plan collision-free trajectories. Trajectory data is stored in a graph data structure within the dataset creation process. Users can access the data using the provided Python script named download_training.py. The dataset supports downloading data for specific cameras like left or right. TartanAir provides a URL list for downloading specific files.
Sources
- https://arxiv.org/html/2003.14338
- https://github.com/castacks/tartanair_tools