This is the pytorch implementation for HDR reconstruction algorithm using deep CNNs, which was proposed in:
Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafał K. Mantiuk, Jonas Unger, "HDR image reconstruction from a single exposure using deep CNNs," ACM Transactions on Graphics, November 2017
I used Kalantari's dataset for training and testing purposes, which was downloaded from:
https://cseweb.ucsd.edu/~viscomp/projects/SIG17HDR/
Since this dataset is used for HDR synthesis task, I only used center images for HDR reconstruction. The LDR and HDR images were split into smaller patches, then an annotation file was created for storing their addresses.
python split_image.py
Specifically, the annotations.txt file was created, which contains the addresses of pairs of LDR-HDR images, for example:
data/train/16-11-04-17/256_512.png data/train/16-11-04-17/256_512.hdr
data/train/16-11-04-17/0_256.png data/train/16-11-04-17/0_256.hdr
data/train/16-11-04-17/768_256.png data/train/16-11-04-17/768_256.hdr
data/train/16-11-04-17/256_0.png data/train/16-11-04-17/256_0.hdr
data/train/16-11-04-17/768_0.png data/train/16-11-04-17/768_0.hdr
data/train/16-11-04-17/512_512.png data/train/16-11-04-17/512_512.hdr
data/train/16-11-04-17/256_256.png data/train/16-11-04-17/256_256.hdr
data/train/16-11-04-17/512_256.png data/train/16-11-04-17/512_256.hdr
data/train/16-11-04-17/768_512.png data/train/16-11-04-17/768_512.hdr
python train.py
Training and validation loss:
python predict.py
Experimental results (For display purpose, gamma correction was applied to HDR images):
