Home

Awesome

MegCup 2022 Team Feedforward

This repository is the official MegEngine implementation of the 3rd place solution (Team Feedforward) in 2022 MegCup RAW image denoising.

Method

We propose a FeedBack-based Restormer (FBRestormer) for lightweight denoising, the number of parameters in this model is smaller than 100K!

<img src='assets/arch.png' width=45% height=45%></img> <img src='assets/SGFM.png' width=45% height=45%></img>

Notice That, the restormer blocks in our architecture are modified by replacing the depth-wise 3x3 convlution in the GDFN with depth-wise 5x5 convlution which is also dilated with dilation equals 2.

The simple gated fusion module is used for feedback connections.

Environment

Conda

$ conda create -f ./env.yaml
$ conda activate megcup

Docker

<font color=red>TBD</font>

Usage

$ python test.py --data-path DATA_PATH      # The test input data path.
                 --checkpoint CHEKPOINT     # The checkpoint need to be loaded.
                [--batch-size BATCH_SIZE]   # OPTIONAL: Batch size for the dataloader,            DEFAULT: 1
                [--num-workers NUM_WORKERS] # OPTIONAL: Number of workers for the dataloader,     DEFAULT: 0
                [--output PATH]             # OPTIONAL: The path to output the final binary file, DEFAULT: '.'

Example:

$ cp PATH/DATA .
$ python test.py --data-path ./DATA --checkpoint ./feedback_restormer.mge

Members

Acknowledgement

This project is based on Restormer, SRFBN, and GMFN.