Single-Image HDR Reconstruction by

Learning to Reverse the Camera Pipeline



Yu-Lun Liu1,5*     Wei-Sheng Lai2*     Yu-Sheng Chen1     Yi-Lung Kao1     Ming-Hsuan Yang2,4     Yung-Yu Chuang1     Jia-Bin Huang3    
1National Taiwan University     2Google     3Virginia Tech     4University of California at Merced     5MediaTek Inc.    

Results on the HDR-Synth, HDR-Real, RAISE, HDR-Eye dataset, and examples from prior work and DxOMark.

We provide the full comparison with all evaluated methods in the paper.

We use the Vibrant tone-mapping mode in Photomatix (https://www.hdrsoft.com/) to generate the tone-mapped HDR images.

The "Ground truth HDR" images in the HDR-Real dataset are generated from fusing multiple LDR images with Photomatix. Therefore, there might have a slight misalignment between the ground truth HDR and input LDR image.

Mouse over the images to flip between the results from our method and others.


Input LDR

Ours*

[Akyüz et al. 2007]

[Huo et al. 2014]

[Kovaleski and Oliveira 2014]

[Masia et al. 2017]

HDRCNN* [Eilertsen et al. 2017]

DrTMO* [Endo et al. 2017]

ExpandNet [Marnerides et al. 2018]

Ground truth HDR