Hi,
The intended goal of this challenge is not clear to me. This is because of two reasons:
1) In the training dataset you have HR images and LR images. I found that HR images are sometimes more blurry than their corresponding LR images. In the first case, If my algorithm produces a blurry result then I am doing the right thing according to the dataset but this is the opposite of what I want from real SR. What do you think about this?
2) Additionally, as was suggested in another post some LR/HR pairs are not aligned properly. For example in image 000395_HR there is a persons head in the image which is not present in 000395_LR.
In this case, my algorithm has to synthesize a head in the HR image but again this is not what I expect from real SR. Maybe the case of the head is an outlier but I found misalignment much more common between LR/HR pairs.
This challenge is stated to be "Real Image Super-Resolution", but from these observations I don't see how the dataset helps us accomplish that.
Posted by: zat @ June 11, 2020, 10:21 a.m.Thanks for your question!
As we all know, existing SISR datasets have a shortcoming of synthetic image degradation, which is limited for realistic applications.
Thus, it deserves our great efforts and attempts for real SR datasets and algorithms. Although building a dataset with more textures and diverse scenes in the real-world is extremely challenging and tedious, we have been carefully doing this work for a long time.
1: strictly control the process of image capture: fix the tripod, use Bluetooth, and take photos in the scenes without moving objects as possible as we can.
2: carefully and iteratively refine the alignment results many times
3: remove images that are not well-aligned and patches that are smooth regions or blurry, and select patches with rich textures for the training set
4: carefully and manually check the alignment results for each image and avoid the misalignment like those you pointed as possible as we can.
5: conduct extensive experiments with existing SR methods to evaluate this dataset
It is an extremely difficult task to exactly align images, especially pixel-wise alignment for SR. Without careful alignment or with rough alignment, SR models are prone to generating blurry results. Thus, we have many efforts to improve the alignment and provide a well-designed real SR dataset as possible as we can. Besides, we have conducted experiments to evaluate this dataset, checked its alignment issue and improved it over and over again. Because we do not hope that some unexpected issues emerge to dominate the task and have significant negative effects when trying to address an important real SR problem. We trained many state-of-the-art SR methods and explored new resolutions on this dataset for different scale factors. When testing those trained models, those extensive experimental results demonstrate a promising performance for real SR and appealing generalization for realistic applications. That is, our efforts for the alignment are reliable and the effect of potential misalignment is very trivia.
Posted by: hannanlu @ June 12, 2020, 9:38 a.m.