Multispectral Imaging (MSI) datasets

In this WWW page five datasets (Dataset 1, Dataset 2, Dataset 3, Dataset 4, Dataset 5) for Multispectral Imaging (MSI), annotated with ground truth data, are availed. These datasets are described and referred to in the following paper:

Athanasios Zacharopoulos, Kostas Hatzigiannakis, Polykarpos Karamaoynas, Vassilis M. Papadakis, Michalis Andrianakis, Kristalia Melessanaki, Xenophon Zabulis, "A method for the registration of spectral images of paintings and its evaluation"

Please cite the above paper, if you use these datasets.

The datasets were acquired using IRIS II which is a lightweight portable system comprising of a high resolution camera, a novel filter wheel able to interchange 23 filter positions and fast electronics and upon the same subject and viewing conditions. The wavelengths that the 23 images of each dataset were acquired are the following: 360nm, 370nm, 380nm, 400nm, 425nm, 475nm, 500nm, 525nm, 550nm, 575nm, 600nm, 625nm, 650nm, 700nm,750nm, 800nm, 850nm, 900nm, 950nm, 1000nm, 1050nm, 1100nm and 1150nm. The µ landmarks establishing ground truth correspondences across all images of the sequence were markers. In particular, µ = 16 small markers were printed using a laser printer on white paper. Each target was a, square, small, black and white 2 x 2 checkerboard, which was clearly distinguishable in each wavelength. The targets were placed in a 4 x 4 grid arrangement covering almost entirely the field of view, as shown in the following figure. Marker size was 25mm2 and markers were imaged by approximately 40x40 pixels. The images are encoded in 16bits, in the PNG image file format. More details about the acquisition of the datasets and the imaged targets can be found in the above paper. More details about the imaged paintings can be found in the above paper.

Sample images from the benchmark Dataset 1 and Dataset 2 for the evaluation of the proposed algorithm, corresponding to the 400, 500, 750 and 1000nm wave-lengths (top to bottom and left to right) are shown in the figures below. Upon the painting, µ = 16 markers can be observed which are arranged in a 4 x 4 grid, for the annotation with ground truth correspondences.

To detect the landmark points accurately, an automatic corner detection algorithm was employed, that would accurately find the location of the checkerboard inner corner, given an initial coarse estimate. More details about this process can be found in the above paper. These centers were set as the landmarks ui,k, comprising the ground truth annotation of the datasets. The order of the centers is consistent in the annotation and, in this way, avails the landmark correspondence information.

Dataset 1 In Dataset 1, imaging distance was ≈ 86.5 cm. FOV covered 190 x 142 mm2. Marker size was 25mm2. The dataset images can be found here. The ground truth annotations can be found here.
The annotations in text files encode the 2D coordinates of marker points in the following format: x y, one coordinate per image line. The annotations are also visualized upon the original images in PDF format and these visualizations can be downloaded here.
Dataset 2
In Dataset 2, imaging distance was ≈ 86.5 cm. FOV covered 190 x 142 mm2. Marker size was 25mm2. The dataset images can be found here. The ground truth annotations can be found here.
The annotations in text files encode the 2D coordinates of marker points in the following format: x y, one coordinate per image line. The annotations are also visualized upon the original images in PDF format and these visualizations can be downloaded here.
Dataset 3
In Dataset 3, the setup of Dataset 1 was replicated but attaching the camera to a tripod, to simulate realistic use cases of on-site acquisition. The dataset images can be found
here. The ground truth annotations can be found here.
The annotations in text files encode the 2D coordinates of marker points in the following format: x y, one coordinate per image line. The annotations are also visualized upon the original images in PDF format and these visualizations can be downloaded here.
Dataset 4
In Dataset 4, the setup of Dataset 3, imaging distance was increased to ≈ 120 cm. Imaged area was 267 x 200 mm2. Marker size was 34.6 mm2. The dataset images can be found
here. The ground truth annotations can be found here.
The annotations in text files encode the 2D coordinates of marker points in the following format: x y, one coordinate per image line. The annotations are also visualized upon the original images in PDF format and these visualizations can be downloaded here.
Dataset 5
In Dataset 5, the setup and markers of Dataset 4 were utilized. A small 3.75o angle was introduced between the camera and the painting. It simulates a typical error in the, ideally perpendicular, camera placement, in cases of on-site image acquisition. The dataset images can be found
here. The ground truth annotations can be found here.
The annotations in text files encode the 2D coordinates of marker points in the following format: x y, one coordinate per image line. The annotations are also visualized upon the original images in PDF format and these visualizations can be downloaded here.


For any issue regarding this page or datasets, please contact:
Xenophon Zabulis
http://www.ics.forth.gr/~zabulis