Calibration RGB-D based on spheres Repository
RFAI - Laboratoire d'Informatique Fondamentale et Appliquée de Tours (EA 6300) - France
The goal of this benchmark is to evaluate RGB-D calibration methods based on spheres using the same dataset. We propose several metrics to evaluate the calibration results, as well as the results provided by three algorithms.
The following details the capture environment, the architecture and the evaluations metrics. For more details, refers to the dataset paper  :
|# RGB-D Pairs||Support||Capture
The source code to perform an RGB-D calibration with these files is not available.
These datasets are taken in the following environment :
In this dataset, images are stored as .PNG, points clouds as .PLY, and configuration files as .YML.
A dataset is organized the same way either it contains real or synthetical data. Linked data have the same name. RGB-D input data, as captured by the RGB-D camera couple is stored in the RGB, POINTS and DEPTH folders. The folder Processed contains pre-computed data, such as the segmented ellipses and point cloud. The folder OUTPUT contains calibration results from several algorithms.
Note that for some dataset, some folders are not provided, either because data was not available, or because it does not make sense (e.g. the folder EllipseImg for synthetical data).
Figure 2: Overview of a dataset organization, and their respective metrics (in light gray) for evaluation of the RGB-D Calibration Process.
In order to evaluate precisely the influence of noise on the calibration result (especially the Extrinsic parameters), a synthetical dataset, with known values, has been constructed. A GroundTruth scene has been made. We applied 3 types of noise, as shown on Figure 4.
1 - Gaussian Noise on the RGB ellipses points
2 - Noise on the RGB camera Intrinsic parameters (we multiply each value by a scalar)
3 - Noise on the Depth data, as a shift of all the point cloud (which displace the point cloud centroid)
As we apply Gaussian Noise, we performed multiple iterations to study the distribution of the resulting values
Figure 3: Overview of a synthetic dataset generation
In the synthetical data, we only modify the Depth Noise between the multiple evaluations (noise n°3), by increasing its amount by 0,4. The noise values are provided in the /noise.yml. All steps are performed a hundred times to have reliable results. Because of the amount of synthetical datas, the Ground Truth scene (/GroundTruth), as well as the results (/results) are provided in separate folders.
To perform a calibration, you can either use the cameras frames (/root), or the ellipse and sphere detection (/Processed). The calibration results have to be stored in the /OUTPUT folder.
To read our multiples files, we recommand the use of the OpenCV library.
We recommend the use of the "cv::RotatedRect" type as output of this file Parameters :
x(double) : x coordinate of the ellipse center
y(double) : y coordinate of the ellipse center
majorAxis(double) : major axis of the ellipse
minorAxis(double) : minor axis of the ellipse
angle(double) : rotation angle (in degree) of the ellipse
x(double) : x coordinate of the sphere center
y(double) : y coordinate of the sphere center
z(double) : z coordinate of the sphere center
u(double) : x coordinate of the projected sphere center onto the depth camera plane (with manufacturer intrinsic parameters)
v(double) : y coordinate of the projected sphere center onto the depth camera plane (with manufacturer intrinsic parameters)
Radius(double) : radius of the fitted sphere
rms(double) : resulting error of the fitting algorithm (Least-Squares)
sphereRadius(double) : radius of the used sphere
relativeSphereDistance(double) : Distance between the two spheres (for Double dataset only)
rgbIntrinsic(Mat3x3[double]): RGB camera intrinsic parameters
depthIntrinsicManufacturer(Mat3x3[double]) : Depth camera intrinsic parameters
rgbDistCoeffs(Mat5x1[double]) : RGB camera distortion coefficients
depthDistCoeffs(Mat5x1[double]) : RGB camera distortion coefficients
fileKeys(array[string]) : File used in the calibration
translation(Mat3x1[double]) : Resulting translation
rotation_degree(Mat3x1[double]) : Resulting rotation
depthIntrinsic(Mat3x3[double]): Resulting Depth camera intrinsic parameters
rgbIntrinsic(Mat3x3[double]) : RGB camera intrinsic parameters
meanReprojectionError(double) : Mean of the reprojection error values
reprojectionError(array[double]): Reprojection error values
This file evaluate the ellipse and sphere detection against the manually segmented ground truth.
index: File name to evaluate
DSC: Sorensen-Dice coefficient for the ellipse detection evaluation (see Equation 2)
L2: L2-norm for the sphere detection evaluation (see Equation 3)
SigmaRGB: Gaussian Noise on the RGB ellipses points (1)
SigmaKr: Noise on the RGB camera Intrinsic parameters (2)
SigmaDepth: Noise on the Depth data (3)
e_r: Reprojection error (see Equation 4)
E_r: 3D point-to-point mean distance (see Equation 5)
R_err: Rigid body rotation error (see Equation 8)
t_err: Rigid body translation error (see Equation 7)
K_diff: Fu_diff + Fv_diff + u0_diff + v0_diff + s_diff (for each *_diff variable, see Equation 6)
Fu_diff: Absolute difference of the focal expressed in pixel width (px)
Fv_diff: Absolute difference of the focal expressed in pixel height (px)
u0_diff: Absolute difference of the u0 coordinate of the principal point (u0, v0) (px)
v0_diff: Absolute difference of the v0 coordinate of the principal point (u0, v0) (px)
s_diff: Absolute difference of the affine parameter (px)
TR_diff: R_diff + t_diff
R_diff: Rx_diff + Ry_diff + Rz_diff
Rx_diff: Absolute difference of the rotation around the x axis (deg.)
Ry_diff: Absolute difference of the rotation around the y axis (deg.)
Rz_diff:Absolute difference of the rotation around the z axis (deg.)
t_diff: tx_diff + ty_diff + tz_diff
tx_diff: Absolute difference of the translation along the x axis (mm)
ty_diff: Absolute difference of the translation along the y axis (mm)
ty_diff: Absolute difference of the translation along the z axis (mm)
 D. J. T. Boas, S. Poltaretskyi, J.-Y. Ramel, J. Chaoui, J. Berhouet, and M. Slimane, "A Benchmark Dataset for RGB-D Sphere-Based Calibration" - PDF
 R. Hartley and A. Zisserman, Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN:0521540518, second ed., 2004
 A. Staranowicz, G. R. Brown, F. Morbidi, and G. L. Mariottini, “Easy-to-Use and Accurate Calibration of RGBD Cameras from Spheres,” in Image and Video Technology, vol. 8333, pp. 265–278, Springer Berlin Heidelberg, 2014
 D. J. T. Boas, S. Poltaretskyi, J.-Y. Ramel, J. Chaoui, J. Berhouet, and M. Slimane, “Relative pose improvement of sphere based rgb-d calibration,” Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, vol. 4, pp. 91–98, 2019
This database is publicly available and free for research staff.
Also, please inform us about your work and, if possible, send a copy of your publication to the contacts mentioned below. We will eventually put a list of references on this web page.
Limitation of Liability :
The images contained in this database are provided 'as is' without warranty of any kind. The entire risk is assumed by the user, and in no event will RFAI be liable for any consequential incidental or direct damages suffered in the course of using the database.
Permission to use but not reproduce or distribute the database is granted to all researchers given that the following steps are properly followed :
Permission is NOT granted to reproduce the database or posted into any other webpage.
None economical profit can be obtained from this database.
Technical questions about the database, the format, or problems obtaining the data should be directed to the database editors:
Lab./Dep .Informatique de Tours - PolytechTours
64, Av. Jean Portalis
37200 TOURS - FRANCE
Tel: +33 188.8.131.52.26