A multi‐illuminant synthetic image test set. Issue 6 (3rd August 2020)

• Record Type: Journal Article
• Authors: Hao, Xiangpeng; Funt, Brian
• Is Part Of: Color research & application. Volume 45:Issue 6(2020:Dec.)
• Page Start: 1055
• Page End: 1066
• Publication Date: 2020-08-03
• Subjects: spectral rendering -- illumination estimation -- color constancy -- full‐spectrum image dataset; Color -- Periodicals; 535
• DOI: 10.1002/col.22535
• Languages: English
• ISSNs: 0361-2317

Abstract

Abstract: A new multi‐illuminant synthetic image test set called MIST is described and made publicly available. MIST is intended primarily for evaluating illumination estimation and color constancy methods, but additional data are provided to make it useful for other computer vision applications as well. MIST addresses the problem found in most existing real‐image datasets, which is that the groundtruth illumination is only measured at a very limited number of locations, despite the fact that illumination tends to vary significantly in almost all scenes. In contrast, MIST provides for each pixel: (a) the percent surface spectral reflectance, (b) the spectrum of the incident illumination, (c) the separate specular and diffuse components of the reflected light, and (d) the depth (ie, camera‐to‐surface distance). The dataset contains 900 stereo pairs, each of the 1800 images being a 30‐band multispectral image covering the visible spectrum from 400 to 695 nm at a 5 nm interval. Standard sRGB versions of the multispectral images are also provided. The images are synthesized by extending the Blender Cycles ray‐tracing renderer. The rendering is done in a way that ensures the images are not only photorealistic, but physically accurate as well.