We borrow the ideas from trichromatic IID benchmark and establish the very first spectral intrinsic ground-truth dataset called SIID dataset. The SIID dataset includes original data without specular variation, which is called diffuse.hsd, and groundtruth data termed as shading.hsd for several groups of objects with different materials such as plastic, glass, cellulose and calcium sulphate dihydrate . Many other vision researches (e.g. IRSS, segmentation, recognition, relighting) can also be benefited by the SIID dataset.
We extract lighting spectra from the scene by exposing the spectral camera directly to the light source. We relax the white light assumption to arbitrary illumination condition and manually set an exposure brightness value to assure that the lighting spectra can be observed with sufficient dynamic range. For each object in our dataset, we employ two kinds of light sources in capturing scene, one is
We use the cross-polarizing approach to separate the diffuse and the specular term and the polarizing filters are placed over both the light source and the spectral camera. Since the light ray reflect off the surface of objects remain polarized, we are able to remove
The scene in our dataset contains non-uniform texture and we remove the texture variation by spraying these objects with a thin layer of flat white oil-painting and rephotograghing it under the same illumination. In this way, we get the
Please use reference “Nanjing University, Lab for Computational Imaging Technology & Engineering” when you use the databases from these pages. This database is made available only for non-commercial use.
Wavelength interval: 410 nm - 700 nm
Wavelength resolution: 5 nm/10nm
Chen X, Zhu W, Zhao Y, Yu Y, Zhou Y, Yue T, Du S, Cao X. "Intrinsic decomposition from a single spectral image". Appl. Opt. 56: 5676-5684, 2017