Dense Image Matching - data description (Dortmund, Zeche Zollern)
RELEASE of ALL THE DATA captured within the ISPRS Scientific Initiative "Multi-platform Very High Resolution Photogrammetry" cofunded by EuroSDR (2014-2015)
All the data acquired for the benchmark has been released on both the test areas. You are welcome to use in any activity where it might be helpful. If you intend to use it in a publication please acknowledge the data provision by ISPRS and EuroSDR. Please also refer to the paper:
A more detailed explanation about the complete dataset will be given by mail after having filled the registration form!
___________________________________________________________________________________________________________
Image orientation performed as a pre-processing step for the dense image matching (DIM) tasks
Bundle block adjustment has been performed with Pix4dmapper, for details see below. Ground control points were measured with static GNSS and a nominal accuracy of sigma=1cm. Targets at the facade got measured with a total station, with an absolute positioning accuracy of the same order of the GNSS acquisition.
Scenario A) Dense image matching in oblique and nadir airborne imagery
The five camera heads in the PentaCam system have been calibrated in a lab and corresponding calibration documents are available from the data download location. Note that the definition of the camera coordinate system as described in the camera calibration document is different from the definition by Pix4Dmapper which we used to perform the aero triangulation. In the download folder you find two different camera parameters: the Pix4D-system and the Match-AT-system. The Match-AT parameters are defined according to the camera calibration document.
The lens distortion has been corrected in a post-processing step by the data provider. however, the original NADIR view still contains significant distortions as described in the calibration report: a folder with undistored nadir images has been therefore added as well. For the oblique views the remaining distortion is - according to the camera vendor - smaller than 2µm, hence we assume distortion free oblique images. If you plan to use also the nadir views for dense image matching make sure you consider the distortion.
The complete image block has been adjusted but only 85 images in correspondence of the "administration building" of the Zeche Museum of Dortmund are delivered in the benchmark. To support the bundle block adjustment and because the used software only matched a few tie points across the viewing directions automatically, some tie points got measured manually. The pix4dmapper report, which is also available in the download folder, indicates the following RMSE values.
- At GCPs: X=1.4cm, Y=0.6cm, Z=14.7cm.Especially one of the three targets at the facade show a larger RMSE in Z, which is because of uncertain identification.
- At Check Points: X=2.1cm, Y=4.5cm, Z=8.3cm
Given the average ground sampling distance of 10cm for the nadir view and 8 to 12 cm for the obliques, the obtained accuracy at the independent check points is acceptable. Image reprojection errors vary around 0.5 pix which is reasonable, as well.
Image (external) orientation information is provided in several formats:
- Pix4D-system
- Traditional photogrammetric format (rotation sequence defined like in photogrammetric workstations like Inpho and LPS): X Y Z omega phi kappa, see file calibrated_external_camera_parameters.txt
- Matrices/vectors per image: Camera matrix K, translation vector t, rotation matrix R, see file calibrated_camera_parameters.txt
- Camera projection matrix P=KR|t, see pmatrix.txt. In order to avoid large numbers in the matrix computation a global offset was applied to the translation vectors shown in the latter two files. See the file README_global_offset.txt and the header of pmatrix.txt.
- MatchAT-system
- The file MATCH_orientations.prj enables to load the whole project with the images in the software MatchAT. Note: the orientation has been performed using the calibration report provided by IGI systems GmbH (see below).
- Complete_Ori folder where all the orientations of the images are reported in different files.
- The orientation are also available in the file orientation_XYZ_opk.dat
Note:
- Although for this task we only release 85 images (just Zeche area), the image orientation files list 881 different images in total (the entire image block of that part of Dortmund). This means that the sequence of images in the image folder is not identical to the sequence in the orientation files. Attention must thus be paid when the images are read!
- The differences between the image camera convention of the Pix4D-system and the MatchAT-system are reported in the file ORIENTATION_explanation.pdf.
The download folder for this Scenario contains the following subfolders:
Calibration_report: calibration report provided by IGI systems GmbH. Note that all images, but the nadir images, have been undistorted in a post processing step, see above. There is also a pdf which describes for each camera the definition of the camera coordinate system in the pix4d definition;
Image_orientation: the files described above, including the pix4d and MatchAT reports (referring to all 881 images);
8bit and 16bit: the 82 images of the respective spectral resolution. The 16bit to 8bit transformation has been done by AeroWest using some batch processing, i.e. no individual adaptations were made. Image matching using the full spectral resolution should theoretically lead to better results. The image file naming scheme is flightline_station_cameranumber000image-of-that-camera.
Scenario B) Dense image matching in a combined terrestrial and UAV image block
In total more than 1000 imagesof the “administration building" of the Zeche Museum have been acquired from a rotary wing platform and the ground. For the image orientation a subset of images got selected (867) and a further reduced subset is released for this benchmark (228 images).
A Sony Nex 7 camera got mounted on the UAV, and for the ground acquisition a Canon D600 SLR camera with 20mm Voigtländer fixed zoom lens was used. In both cases the focus was fixed at infinity, hence we assume a sufficiently constant interior orientation for both cameras for the duration of image acquisition.
The interior parameters got estimated in a self-calibration fashion, i.e. as additional unknowns during bundle adjustment. For the data release the images got undistorted by pix4d applying the estimated distortion parameters. From this step, however, no quality indication, i.e. remaining distortions, is known. Some sample test where straight lines got fitted to straight building edges did not reveal significant distortion.
The pix4dmapper report, which is also available in the download folder, indicates the following RMSE values:
- GCPs: X=0.5cm, Y=0.6cm, Z=0.8cm.
- Check Points: X=0.5cm, Y=0.7cm, Z=0.1cm
The average GSD is 0.5cm. The remaining errors are in fact smaller than the nominal accuracy delivered by the GNSS and total station measurements.
Image orientation and calibration information is provided in several file formats. In addition to the formats listed for Scenario A, we have here the interior camera parameters as computed within the self-calibration approach, see *.cam-files in the folder image_orientation. Camera id=0 refers to the Canon 600D (terrestrial shots), while id=1 refers to the Sony Nex 7, mounted on the UAV. Note that the images got undistorted already, hence, only the principal point location and focal length needs to be considered. Also note that the pixel aspect ratio is constrained to 1. The image folder contains the 228 undistorted images. Image names containing the string IMG are from the Canon, while the DSC-images are captured by the Sony device.
TERM OF USE
The data users should include the following acknowledgement in any publication resulting from the dataset:
"The authors would like to acknowledge the provision of the datasets by ISPRS and EuroSDR, released in conjuction with the ISPRS scientific initiative 2014 and 2015, lead by ISPRS ICWG I/Vb."
ICWG II/Ia
