Xiangpeng Hao
2020 Winner
Simon Fraser University
A Multi-illuminant synthetic image test set
​
Until relatively recently, the majority of work on colour constancy has been based on the assumption that the scene is lit by a single illuminant. Everyone is aware that this assumption is very often violated, but there are only very limited image datasets providing groundtruth data for multi-illuminant scenes. Furthermore, these data sets are small and mainly consist of images of quite constrained single-object scenes.
​
In response, we have developed and made public a new image test set called MIST that includes 900 full-spectrum stereo image pairs of complex scenes. The images are synthetically generated and accompanied by pixelwise groundtruth data that includes, for each pixel, its percent surface spectral reflectance, the spectrum of the incident illumination and its depth (i.e., camera to surface distance). The pixel wise spectrum of the reflected light is also provided in terms of its separate diffuse and specular components. Although the primary objective in constructing MIST is to aid in the evaluation of illumination estimation methods for colour constancy, the inclusion of depth and stereo data is expected to be useful for testing other computer vision and hyper spectral algorithms as well.