[IROS2020] Author-collected real-world datasets

paper link : https://ieeexplore.ieee.org/abstract/document/9341486

The datasets are used in our submission to IROS 2020. C. Kim, J. Kim and H. J. Kim, “Edge-based Visual Odometry with Stereo Cameras using Multiple Oriented Quadtrees,” 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 5917-5924.

Please use this to cite: @INPROCEEDINGS{kim2020edgebased, author={Kim, Changhyeon and Kim, Junha and Kim, H. Jin}, booktitle={2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, title={Edge-based Visual Odometry with Stereo Cameras using Multiple Oriented Quadtrees}, year={2020}, volume={}, number={}, pages={5917-5924}, doi={10.1109/IROS45743.2020.9341486}}

Characteristics

The datasets include images taken in long and looped corridor, empty room in a modern building.

All scenes are obtained in challenging circumstances with no points and straight lines. In addition, scenes undergo drastic illumination changes in several time intervals.

How to use

After reading details below, you can download each dataset at the Download List section below.

For convenient employing, we provide development tools in MATLAB environments, but you also can use the datasets in other environments by your own parsing schemes.

File Formats

We provide time-stamped stereo images as .zip and .tar files (ZIP and TGZ). Each dataset contains:

  • two stereo images as grayscale 480 X 640 8-bits images in PNG format at 20 Hz.
  • Calibration data (intrinsic and stereo extrinsic parameters) as a form of yaml files.
  • Association file containing timestamps and file names for each stereo image pair as a txt file.

Stereo Imgaes

We use a stereo camera with 12 cm baseline at 20 Hz produced by withrobot co. provider site. All images are accurately time-synchronized by a hardware internal clock of the oCamS stereo camera.

  • Resolution: 480 X 640 (8-bits monochrome) x 2 (left and right)
  • File extension: PNG
  • Frequency: 20 Hz

Association File

To load the stereo images, we provide association files in txt format.

  • Each line in the text file contains synchronized timestamp and image (File names of both left and right are same.).
  • The file may contain comments which begin with “#”.

Intrinsic and Extrinsic Parameters

Although we provide the both parameters of the stereo camera in each dataset, values are shown below. All datasets share same parameters.

# left intrinsic
fx = 483.601048 # focal length x
fy = 482.859102 # focal length y
cx = 328.272852 # optical center x
cy = 244.932776 # optical center y

# left radial distortion coefficient
k1 = -0.413868
k2 = 0.153386

# left tangential distortion coefficient
p1 = -0.000386
p2 = 0.000112

# right intrinsic
fx = 483.963040 # focal length x
fy = 483.291119 # focal length y
cx = 322.556621 # optical center x
cy = 248.655907 # optical center y

# right radial distortion coefficient
k1 = -0.413974
k2 = 0.153772

# right tangential distortion coefficient
p1 = -0.000077
p2 = -0.000176

# Rotation matrix from left camera to right camera represented in left camera coordinate frame.
R_lr = [0.999941, -0.004345, 0.009982,
0.004335, 0.999990, 0.000979,
-0.009986, -0.000936, 0.999950]

# Translation vector from left camera to right camera represented in left camera coordinate frame.
t_lr = [0.120237, -0.000416, 0.000035]

The above values are achieved by Kalibr program in github.

Ground-truth

Even though there is no ground truth of the datasets, you can qualitatively validate the VO and 3-D reconstruction performance by below rules:

  1. table_and_chairs: Shapes of the upper and bottom board of the table are coincident circles. You can validate your own results in this point.

  2. white_chair: All the shapes of parts are curved.

  3. room_loop: At both first and last images, an identical 9 x 8 (0.07 x 0.07 m squares) checkerboard is observed. You can get a relative 3-D positions and poses of the start and edge images, and use to calculate a final distance error of the estimated trajectory. Also, the camera motion along the height direction is maintained consistently.

  4. corridor_loop: You can compare the start and end points of the estimated trajectories. We carefully move the camera to make both start and end points of the trajectory. Furthermore, the height throughout the trajectory is maintained consistently.

Download List

  • We will update rosbag format.
  • Total lengthes are estimated by our VO algorithm.
  • The datasets can be easily loaded and plotted in MATLAB environment using dataset_tools provided by us. It also contains a rectification tool.
sequence name duration [s] # of images total length [m] avg. vel [m/s]
dataset_tools
ZIP(256KB), TAR(365KB)		 - - - -
table_and_chairs
ZIP(0.06GB), TAR(0.17GB)		 14.45 s 289 6.539 m 0.253 m/s
white_chair
ZIP(0.11GB), TAR(0.30GB)		 25.80 s 516 6.037 m 0.418 m/s
room_loop
ZIP(0.35GB), TAR(0.92GB)		 79.70 s 1,594 26.532 m 0.333 m/s
corridor_loop
ZIP(0.47GB), TAR(1.17GB)		 102.35 s 2,047 46.63 m 0.466 m/s
  • For questions, comments or suggestions, please feel free to send me an e-mail.