search

yjsunnn / FBANet

Official implementation of ICCV2023 "Towards Real-World Burst Image Super-Resolution: Benchmark and Method"
MIT License
113 stars 7 forks source link

Further Information Request About The RAW Data #10

Closed JiangtianPan closed 9 months ago

JiangtianPan commented 9 months ago

Hi, I am excited about this work's result, and I appreciate your open source. May I wonder how to use the "Raw Data" in the dataset?

Firstly, the data are all PNG files, rather than binary files. Secondly, the information obtained from the ".pkl" file is also confusing, without a script, users may not be able to use them to train or test the model.

Could you please exhibit a script to preprocess the raw dataset that can be used to feed the model?

Many thanks for your time.

yjsunnn commented 9 months ago

Sure. The original file of raw processing is on the remote server which has collapsed for a long time, that's why we haven't uploaded an official version of raw data. But I have some temporary files on my own laptop, and I think it will help! The files could be found in the [baidu link]. It contains the data processing part, training part, evaluation part (in the test part of train.py), etc.

yjsunnn commented 9 months ago

And I think the following answer from [another issue] gonna help.

For the raw model, actually there are two ways you can choose: 1) Import the 4-channel data (80x80x4), flatten it (160x160x1) and use the default demosaic function in colour_demosaicing to demosaic it (160x160x3). The rest is quite the same as in RGB. (Our choice); Or 2) Import the 4-channel data, change the first convolution (as you have said, from 3 to 4) and add one more upsampler layer (A popular choice). You can try both ways, and welcome to discuss the results!

yjsunnn commented 9 months ago

Please feel free to re-open the issue if you have any problem.