Abstract
The dataset contains 44 hyperspectral near-infrared images of bulky waste samples. For each hyperspectral image, there is also an RGB and a label image assigning one of 9 classes to each pixel. The majority of samples are crushed furniture from IKEA and the other objects are from a real waste plant or self-collected bulky waste. Details of the file structure are given in "readme.txt" of the provided dataset. The measurements are part of the work described in https://publica.fraunhofer.de/entities/publication/dde5f89c-f87b-4644-b084-f4edf66c4b19/details
Technical Information
Three tiff-files files per sample collection are given in this dataset, with 44 bulky waste sample collections in total. One RGB image (3300x4782 pixel), one hyperspectral image (550x797 pixel with 210 channels) and one gray-scale image (3300x4782 pixel) encoding a class in each pixel (class index = gray scale value / 12). For hyperspectral imaging, the camera FX17e from SPECIM was chosen. The camera collects hyperspectral images with 224 bands ranging from 900nm to 1700 nm. 14 channels at the edge of the spectrum have been removed in the provided images, due to a low sensitivity. The frame rate was set to 104.17 Hz, resulting in a resolution of 1mm/pixel in both axes of the image. A prism-based RGB line scan camera (SW-4000T-10GE) was used to make recordings of the sample scenario. These images were utilized for labelling and conventional RGB image analysis. The frame rate of the RGB camera was set to 625 Hz. The spatial resolution was 0.15mm/pixel. As a light source, halogen lamps were used for both cameras. By moving the samples on a conveyor belt with a speed of 0.108 m/s, images with two spatial axes were constructed using the push-broom method. The scanning lines of the hyperspectral and the RGB camera differ by a few centimeters, which was corrected using a marker-based registration. Each pixel in a hyperspectral image corresponds to a 6x6 pixel region in the RGB and the label image. To have a one-to-one pixel correspondence for all images, one can downsample the RGB and label images by a factor of 6.
