Latest revision as of 20:20, 1 February 2021

Abstract

The evaluation of computer vision methods on synthetic images offers control over scene, object, and camera properties. The disadvantage is that synthetic data usually lack many of the effects of real cameras that pose the actual challenge to the methods under investigation. Among those, noise is one of the effects more difficult to simulate as it changes the signal at an early stage and is strongly influenced by the camera’s internal processing chain. The resulting noise is highly complex, intensity dependent, as well as spatially and spectrally correlated. We propose to transform synthetic images into the raw format of digital cameras, alter them with a physically motivated noise model, and then apply a processing chain that resembles a digital camera. Experiments show that the resulting noise exhibits a strong similarity to noise in real digital images, which further decreases the gap between synthesized images and real photographs.

Original document

The different versions of the original document can be found in:

• https://elib.dlr.de/127803

• https://elib.dlr.de/127803/1/Bielova_A_Digital_Image_Processing_Pipeline_for_Modelling_of_Realistic_Noise_CVPRW_2019_paper.pdf

• http://xplorestaging.ieee.org/ielx7/8972688/9025328/09025558.pdf?arnumber=9025558,

http://dx.doi.org/10.1109/cvprw.2019.00350

• https://dblp.uni-trier.de/db/conf/cvpr/cvprw2019.html#BielovaHLH19,

https://openaccess.thecvf.com/content_CVPRW_2019/papers/PCV/Bielova_A_Digital_Image_Processing_Pipeline_for_Modelling_of_Realistic_Noise_CVPRW_2019_paper.pdf,

http://openaccess.thecvf.com/content_CVPRW_2019/html/PCV/Bielova_A_Digital_Image_Processing_Pipeline_for_Modelling_of_Realistic_Noise_CVPRW_2019_paper.html,

https://elib.dlr.de/127803,

https://academic.microsoft.com/#/detail/2972870201