Residual Networks for Light Field Image Super-Resolution
CONTACT: Shuo Zhang
(zhangshuo@bjtu.edu.cn)
Any scientific work that makes use of our code should appropriately mention this in the text and cite our CVPR 2019 paper. For commercial use, please contact us.
Shuo Zhang, Youfang Lin, Hao Sheng.
Residual Networks for Light Field Image Super-Resolution,
IEEE Conference of Computer Vision and Pattern Recognition, 2019
Test example:
For central view:
python resLF_test.py -I image_test_path/ -M model_path/ -S save_path/ -o 9 -c 3 -g 0 -s 2 -i blur -C y
For full Light Field SR (7*7):
python resLF_test.py -I image_test_path/ -M model_path/ -S save_path/ -o 9 -c 7 -g 0 -s 2 -i blur -C n
For Help
python resLF_test.py -h
Train example:
We provide different downsampling methods named 'Bicubic' and 'Blur'. In 'Blur', the sub-aperture images are first blurred using the normalized box filter (the filter size is $scale \times scale$), then regularly decimated to the desired resolution. In 'Bicubic', the sub-aperture images are directly downsampled using bicubic downsampling method.
Note that 'Blur' downsampling method achieves better results compared with 'Bicubic' downsampling method.
The training and testing example dataset can be found at: http://lightfields.stanford.edu/ (Lytro Illum), https://mmspg.epfl.ch/downloads/epfl-light-field-image-dataset/ (Lytro Illum) and https://lightfield-analysis.uni-konstanz.de (Synthesis Light Field). See Training/Testing Dataset.txt.
In order to test the Miscellaneous Category
dataset in http://lightfields.stanford.edu/, we retrain the model with a new training dataset. (Our training dataset in CVPR paper includeds the Miscellaneous Category
.) The new list and the trained model is provided, in which we als obtain better performances compared with the results in the CVPR paper.
python 3.6.5
pytorch 0.4.0
cuda 10.1
cudnn 7.5.1
2019.10.28 4X models are provided.
2019.09.19 Different downsampling methods are provided.
2019.06.12 The package released.