From Pixels to Physics: Probabilistic Color De-rendering

TitleFrom Pixels to Physics: Probabilistic Color De-rendering
Publication TypeConference Paper
Year of Publication2012
AuthorsXiong, Y., Saenko K., Darrell T., & Zickler T.
Other Numbers3451
Abstract

Consumer digital cameras use tone-mapping to producecompact, narrow-gamut images that are nonetheless visuallypleasing. In doing so, they discard or distort substantialradiometric signal that could otherwise be used for computervision. Existing methods attempt to undo these effectsthrough deterministic maps that de-render the reportednarrow-gamut colors back to their original wide-gamut sensormeasurements. Deterministic approaches are unreliable,however, because the reverse narrow-to-wide mappingis one-to-many and has inherent uncertainty. Our solutionis to use probabilistic maps, providing uncertaintyestimates useful to many applications. We use a nonparametricBayesian regression technique—local Gaussianprocess regression—to learn for each pixel’s narrow-gamutcolor a probability distribution over the scene colors thatcould have created it. Using a variety of consumer cameraswe show that these distributions, once learned from trainingdata, are effective in simple probabilistic adaptations oftwo popular applications: multi-exposure imaging and photometricstereo. Our results on these applications are betterthan those of corresponding deterministic approaches, especiallyfor saturated and out-of-gamut colors.

Acknowledgment

This work was partially supported by funding provided to ICSI by the U.S. Defense Advanced Research Projects Agency (DARPA). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors or originators and do not necessarily reflect the views of DARPA or of the U.S. Government. This work was also supported by Toyota and Google, as well as funding provided to ICSI through National Science Foundation grants IIS-0905647 and IIS-0819984. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors or originators and do not necessarily reflect the views of the National Science Foundation.

URLhttps://www.icsi.berkeley.edu/pubs/vision/ICSI_pixelstophysics12.pdf
Bibliographic Notes

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Providence, Rhode Island

Abbreviated Authors

Y. Xiong, K. Saenko, T. Darrell, and T. Zickler

ICSI Research Group

Vision

ICSI Publication Type

Article in conference proceedings