Framework for Illumination Estimation and Segmentation in Multi-Illuminant Scenes

Color constancy is an important part of the human visual system, as it allows us to perceive colors objects invariant color illumination that illuminating them. Modern digital cameras have be able recreate this property computationally. However, not a simple task, response each pixel on camera sensor product combination spectral characteristics illumination, object, and sensor. Therefore, many assumptions made approximately solve problem. One common procedure was assume only one global source illumination. assumption often broken in real-world scenes. Thus, multi-illuminant estimation segmentation still mostly unsolved In paper, we address problem by proposing novel framework capable estimating per-pixel any scene with two sources The consists deep-learning model segmenting image into regions uniform models single-illuminant estimation. First, produced, used input along original image, which segments where illuminant dominant. output mask masked images are given models, produce final illuminations. comprising first trained separately, then combined fine-tuned jointly. This utilize well researched scenario. We show such approach improves both capabilities. tested different configurations proposed against other single- large dataset images. On dataset, achieves best results, multi-illumination problems. Furthermore, generalization properties were datasets. There, achieved comparable performance state-of-the-art single-illumination even though