Team Objective
Move along PSNR - Per curve
1. Undergrad's tips and tricks:
-
- [ ] Mean shift per patch
-
- [ ] RGB shuffle
2. DownMSE loss
3. Architecture adjustment
4. Laplace form
5. Handle faces directly
Overall Progress
08/21 Tuesday:
A. RGBshuffle
-
- [x] Fix directory saving for model testing
-
- [x] Implement rgb random shuffle, random in channels
-
- [x] Test & confirm the shuffling
-
- [x] save the shuffle img patches for sanity check
- pre-shuffle & post-shuffle for both lr, hr
-
- [x] Check the rgb training results, compare with the previous results
-
- [ ] understand the function __get_patch()
-
- [ ] Analysis and understand RGB results difference in paper
Mean shift per patch
B. DownMSE loss
-
- [ ] Save the LR images from SR, outside program for sanity check
- Decide to use BiCubic Downsampling
-
- [ ] Integrate to program, bicubic down-sampling
-
- [ ] Implement downMSE
C. others:
-
- [ ] Display previous running results F_per vs. PSNR score
-
- [ ] Reply Cynthia
-
- [ ] Draft office desk email
Readings & theoretical analysis
-
- [ ] Read about PSNR theory
-
- [ ] Analysis GAN loss in different scalers
- save and load the approximator
-
PDSR Prior work
Submission to the PIRM2018 Super Resolution Contest
The models are to big to fit on github, so they can be found here: https://drive.google.com/drive/folders/13eSV6d5cIFG67MMZ1by54uJcaExd6FHK?usp=sharing