Reimplementation of resLF(CVPR2019): Residual Networks for Light Field Image Super-Resolution (http://openaccess.thecvf.com/content_CVPR_2019/papers/Zhang_Residual_Networks_for_Light_Field_Image_Super-Resolution_CVPR_2019_paper.pdf)
The official code released by authors is at (https://github.com/shuozh/resLF), which is not testable because of the unavailable test images.
I notice that in the official code the test images are in the png format. However, most avaiable light field(LF) test images are in the mat format.
On the one hand, we transform the LF data into png format and have not change any code in resLF_test.py script.
Please download the pre-trained model in the folder.
python resLF_test.py -I 'data_png/' -M 'model/' -S save_path/ -o 9 -c 7 -g 0 -s 2 -i blur -C n
python resLF_test.py -I 'data_png/' -M 'model/' -S save_path/ -o 9 -c 7 -g 0 -s 2 -i bicubic -C n
bicubic x2 | Avg. | max. | min. |
---|---|---|---|
MonasRoom | 41.17/0.9888 | 42.35/0.9918 | 39.15/0.9819 |
Buddha | 39.57/0.9836 | 40.98/0.9877 | 38.46/0.9784 |
blur x2 | Avg. | max. | min. |
---|---|---|---|
MonasRoom | 39.37/0.9816 | 40.76/0.9870 | 38.08/0.9753 |
Buddha | 37.54/0.9739 | 39.03/0.9801 | 36.03/0.9653 |
However, the outputs we get contain artifacts and do not reflect the results in the paper
On the other hand, we downsample the test images(in the mat format) in Matlab to get the low-resolution inputs and write our own script for testing.
Please download the pre-trained model in the folder. The test inputs can be generated by 'generate_lr.m' script.
python eval.py -image_path 'data_mat/' --model 'model/' --scale 2 --view_n 7 --interpolation bicubic --gpu_no 0
We only list the results of bicubic interpolation (x2) on the Buddha and Mona.
Name | Avg | Max | Min |
---|---|---|---|
Buddha | 39.33/0.9825 | 40.66/0.9866 | 38.29/0.9744 |
MonasRoom | 40.89/0.9879 | 41.93/0.9907 | 38.96/0.9809 |
We also made some attempts to achieve better results and got the following performance by clipping 7 edge pixels for testing
Name | Avg | Max | Min |
---|---|---|---|
Buddha | 39.82/0.9825 | 40.94/0.9866 | 38.71/0.9744 |
MonasRoom | 41.19/0.9879 | 42.20/0.9907 | 39.30/0.9809 |
However, there is still a gap between the results and those in the paper
Try to train the resLF from scratch.