FelixOtt/industry_datasets

Industry Datasets

Latest update: 2023-04-19

Project size: 798 GB

Introduction

We publish the Industry datasets that contain a large variety of different image-based and inertial-based dataset recorded at the Fraunhofer IIS test and application center in Nürnberg, Germany. The recording environment is a large indoor industrial hall with three warehouse racks and absorber wall. We record different scenarios, either on a forklift truck, on a positioning system, handheld, or on a small robotic system. Along the images, we provide the ground truth trajectories (3D positions and orientations as quaternions) recorded with an infrared system or a Qualisys system, and the inertial data. The goal is to predict the absolute pose (position and orientation) of the forklift, robot, or human based on images, or the relative pose based on visual and/or inertial odometry.

Datasets

For the Industry Scenario #1 dataset (the Warehouse dataset), see https://www.iis.fraunhofer.de/de/ff/lv/lok/opt1/warehouse.html

The ground truth files are included as a .txt file in each dataset folder and contains the image file names and the ground truth pose (3d position and 4d orientation as quaternion) in the following order:

[path_images]/[image_name].png x_pos y_pos z_pos rot_w rot_p rot_q rot_r

While we train and test scenario #2 and scenario #3 separately, each train/test dataset of the scenario #4 dataset can be cross-validated. The image sizes are 1.280 x 720 pixels. Note that the datasets are recorded with different camera angels (dependent on the scenario setup). Note that partly some images are included that do not have a corresponding ground truth pose (i.e., at the beggining and at the end of a trajectory). Please load only the images from the ground_truth.txt files. For the Industry scenario #4 dataset (the IndustryVI dataset), we additionally recorded the inertial data combined for all datasets in the folder "inertial_data.zip", each including the files imu.csv, magnetic_field.csv, and reference_pose.csv. Note that some datasets are recorded in two separate trajectories (training and test set combined), separated in each folder into recording0 and recording1. Both recordings should be combined and then separated into a 80%/20% training/validation split. Dataset trajectory plots are shown in the folder plots_dataset_images. Result plots are shown in the folder plots_results. We use structure for motion to extract a point cloud of the scenario (see section "structure for motion"). In addition, we generate synthetic images of the environment (see section "synthetic images and point clouds").

Scenario	Dataset	Description	Number images	Data size
Industry scenario #1	train/test	Recorded on positioning sytstem and forklift truck, eight 67° cameras, train: 2 traj., test: 8 traj.	464,804	31.8 GB
Industry scenario #2	train/test	Recorded on positioning system, three 170° cameras	14,955	3.6 GB
Industry scenario #3	train/test	Recorded on a forklift truck, four 170° cameras	5,853	2.0 GB
Industry scenario #4	train1	Recorded on a turtle bot, one camera: clear environment	92,668	30 GB
Industry scenario #4	train2	Recorded on a turtle bot, one camera: four absorber walls with objects	19,872	6.4 GB
Industry scenario #4	train3	Recorded on a turtle bot, one camera: four absorber walls	138,233	42 GB
Industry scenario #4	train4	Recorded handheld, person 1, one camera: four absorber walls	27,856	8.5 GB
Industry scenario #4	train5	Recorded handheld, person 2, one camera: four absorber walls	28,113	9.2 GB
Industry scenario #4	train6	Recorded on a turtle bot, one camera: open environment with objects	114,998	38 GB
Industry scenario #4	train7	Recorded on a turtle bot, one camera: open environment with objects and labyrinth	29,240	9.7 GB
Industry scenario #4	train8	Recorded on a turtle bot, one camera: open environment with objects, labyringth, and absorber walls	110,923	34 GB
Industry scenario #4	test1	Recorded on a turtle bot, one camera: clear environment	23,168	7.5 GB
Industry scenario #4	test2	Recorded on a turtle bot, one camera: one absorber wall	30,379	9.9 GB
Industry scenario #4	test3	Recorded on a turtle bot, one camera: two absorber walls	24,752	7.7 GB
Industry scenario #4	test4	Recorded on a turtle bot, one camera: three absorber walls	26,877	8.3 GB
Industry scenario #4	test5	Recorded on a turtle bot, one camera: four absorber walls with objects	4,968	1.6 GB
Industry scenario #4	test6	Recorded on a turtle bot, one camera: four absorber walls	34,559	11 GB
Industry scenario #4	test7	Recorded handheld, person 1, one camera: four absorber walls	6,964	2.3 GB
Industry scenario #4	test8	Recorded handheld, person 2, one camera: four absorber walls	7,029	2.4 GB
Industry scenario #4	test9	Recorded on a turtle bot, one camera: open environment with objects	28,750	9.4 GB
Industry scenario #4	test10	Recorded on a turtle bot, one camera: open environment with objects and labyrinth	7,311	2.3 GB
Industry scenario #4	test11	Recorded on a turtle bot, one camera: open environment with objects, labyringth, and absorber walls	27,732	8.5 GB
Inertial (scenario #4)	train/test	IMU and magnetic field recordings of handheld recordings and recordings with the turtle bot	-	0.7 GB
Industry synthetic	train/test	Generation of synthetic images from a point cloud (recorded with a NavVis M4 system)	319,955	103 GB

Structure for Motion (SfM)

For SfM, we use an image subset (the image names and the corresponding ground truth poses are in the folder sfm_images). The checkerboard images are in the folder sfm_checkerboard_images. We use the following camera calibration matrix for methods based on SfM for the Industry scenario #4 datasets:

[548.44934818 0.0 317.73762648]
[0.0 549.17600512 249.00614224]
[0.0 0.0 1.0]

Source code will be updated soon (in the folder sfm_code) and will include:

• Structure for motion with hyperparameter searches
• PoseNet for absolute pose regression
• PoseNet for relative pose regression
• Combination of SfM (absolute poses) and PoseNet (relative poses) with pose graph optimization (PGO) to improve the absolute poses

Trajectories

Industry Scenario #2

The first two images are exemplary images of the environment. The third and fourth image show the training and test trajectories of the positioning system.

Industry Scenario #3:

The first image shows the recording setup of the forklift truck. The second image is an exemplary image of a camera attached to the forklift truck. The third and fourth image show the training and test trajectories of the positioning system.

Industry Scenario #4:

Note that the Industry scenario #4 train 4, train 5, test7, and test8 are recorded handheld. The remaining datasets are recorded in a small robotic system near the floor.

Synthetic Images and Point Clouds

We recorded with a NavVis M6 system images of the industrial environment. Here, we recorded two setups of an open environment (see file navvis_open_environment.zip) and with absorber walls (see file navvis_absorber_walls.zip). We use the absorber wall recording (see first figure below) to generate synthetic images. Next, we create a point cloud in a simulator (see Figure 2 to 6 below), and create synthetic images with a fisheye field of view with three camera directions (front, left, right). These datasets are available in the folder synthetic_datasets with the corresponding ground truth absolute poses (3D position and 4D orientation). We record 5 training and 8 testing trajectories (see the plots in the folder synthetic_datasets/trajectory_plots/). All 5 training sets combined is the dataset that we use for training. The ground truth poses are in the corresponding folder.

Trajectories:

References

For more information of the datasets and results, see your publications. If you use our datasets for your research, please consider citing: Link Paper 1, Link Paper 2

@inproceedings{ott_cvprw2020,
  author = {Felix Ott and Tobias Feigl and Christoffer Löffler and Christopher Mutschler},
  title = {{ViPR: Visual-Odometry-aided Pose Regression for 6DoF Camera Localization}},
  booktitle = {\href{https://ieeexplore.ieee.org/document/9150733}{Proc. of the IEEE/CVF Intl. Conf. on Computer Vision and Pattern Recognition Workshops (CVPRW)}},
  pages = {187--198},
  month = jun,
  year = {2020},
  address = {Seattle, WA},
  doi = {10.1109/CVPRW50498.2020.00029}
}

@inproceedings{ott_raichur_arxiv2022,
  author = {Felix Ott and Nisha Lakshmana Raichur and David Rügamer and Tobias Feigl and Heiko Neumann and Bernd Bischl and Christopher Mutschler},
  title = {{Benchmarking Visual-Inertial Deep Multimodal Fusion for Relative Pose Regression and Odometry-aided Absolute Pose Regression}},
  booktitle = {\href{https://arxiv.org/abs/2208.00919}{arXiv preprint arXiv:2208.00919}},
  month = aug,
  year = {2022},
  doi = {10.48550/arXiv.2208.00919}
}

@inproceedings{ott_arxiv2023,
  author = {Felix Ott and Lucas Heublein and David Rügamer and Bernd Bischl and Christopher Mutschler},
  title = {{Fusing Structure from Motion and Simulation-Augmented Pose Regression from Optical Flow for Challenging Indoor Environments}},
  booktitle = {\href{https://arxiv.org/abs/2304.07250}{arXiv preprint arXiv:2304.07250}},
  month = apr,
  year = {2023},
  doi = {10.48550/arXiv.2304.07250}
}

Acknowledgment

This work was supported by the Federal Ministry of Education and Research (BMBF) of Germany by the Bavarian Ministry for Economic Affairs, Infrastructure, Transport and Technology through the Center for Analytics-Data-Applications (ADA-Center) within the framework of "BAYERN DIGITAL II".

License

This work is licensed under a Creative Commons Attribution 4.0 International License, see http://creativecommons.org/licenses/by/4.0/

Kontakt

If you have any questions or tips to improve the datasets, contact me:

Felix Ott: felix.ott@iis.fraunhofer.de

Nordostpark 84, 90411 Nürnberg, Germany, GoogleMaps

Fraunhofer IIS Nürnberg