dataset-processing is a set of scripts to process, convert and visualize the Rosario dataset

Disclaimer

This site and the code provided here are under active development. Even though we try to only release working high quality code, this version might still contain some issues. Please use it with caution.

Ground-truth

The Ground-Truth is generated directly with the GPS-RTK positional measurements. To this end firstly the NMEA are converted to the format timestamp x, y, z using

python <dataset-processing>/nmea2gps.py --input <gps_log> -output <rosatio_gt>

Then, this format could be converted to any desired standard format as TUM or KITTI using the respectively converter:

python rosario_gt2tum.py -g <rosatio_gt> -o <rosario_gt_kitti>
python rosario_gt2kitti.py -g <rosatio_gt> -o <rosario_gt_tum>

Visualization tools

GPS visualization with mapviz

For the visualization of GPS measurements on ROS we recomend mapviz

roscore &
roslaunch <dataset-processing>/visualize_navsatfix.launch
rosbag play <rosbag>

Tested SLAM systems

robot_localization

roscore &
rosparam set use_sim_time true
roslaunch <dataset-processing>/robot_localization/dual_ekf_navsat.launch
rosbag play --clock <sequenceXX.bag>

S-PTAM

roscore &
rosparam set use_sim_time true
roslaunch <dataset-processing>/slam_configs/sptam_zed_bag.launch
rosbag play --clock <sequenceXX.bag>

ORB-SLAM2

roscore &
rosparam set use_sim_time true
rosbag play --clock sequence01.bag /stereo/left/image_raw:=/camera/left/image_raw /stereo/right/image_raw:=/camera/right/image_raw
rosrun ORB_SLAM2 Stereo ~/catkin_ws/src/ORB_SLAM2/Vocabulary/ORBvoc.txt <dataset-processing>/slam_configs/orbslam_ros.yaml true
rosbag play --clock <sequenceXX.bag>

Visual-Inertial Odometry Systems (NEW)

We suggest to take a look at our new publication:

J. Cremona, R. Comelli, T. Pire, Experimental evaluation of Visual-Inertial Odometry systems for arable farming. Journal of Field Robotics, 2022. DOI: https://doi.org/10.1002/rob.22099

There we have tested the most relevant state-of-the-art Visual-Inertial Odometry Systems including ORB-SLAM3, Basalt, Kimera-VIO, etc. This is the corresponding repository and the submitted version of the paper can be found here.

Dataset evaluation

We recomend the evo evaluation tool.

Use the python 2.7 version:

sudo pip2.7 install evo --upgrade --no-binary evo

Examples:

evo_traj tum --ref=gt_tum_02.txt FrameTrajectory_TUM_Format.txt -p --plot_mode=xyz --align

Convert raw data to bagfiles and ASL format

Install dependencies

sudo apt install pv

Apart from this, it is needed Python 3 to run these scripts.

Run convertion script

A bash script is provided to generate rosbag files from raw data. In addition, a tar containing files as ASL format is also generated for each sequence. To do this, after downloading, joining and extracting raw data, run:

./compressAndSplit.sh -dea

Flags (see below for more details):

-d: drop last part of sequences 01 and 06.
-e: fix timestamp issue by making them equidistant.
-a: generate ASL format.

Raw files and generated files can be found in the official Rosario dataset site.

We have experienced issues when collecting data with the IMU we have used to record the dataset. After approximately 458 seconds of recording, erroneous measurements are obtained for a few seconds. We performed several experiments to check this hardware issue.

Duration of sequences 01 and 06 is greater than 458 seconds. We provide a -d flag to cut sequences 01 and 06, discarding the last part of these sequences. Duration of the generated rosbags is approximately 445 seconds (this value can be easily edited in compressAndSplit.sh).

Additionally, an issue related to the timestamps was fixed. We found that multiple rows were labeled with the same timestamp. To solve this, -e flag allows to fix timestamps making them equidistant. See modify_bag_imu_timestamps.py for more details.

This has been better explained in the aforementioned new publication.

CIFASIS / dataset-processing

readme