Official Code for the paper TransWeather, Arxiv Tech Report 2021
About this repo:
This repo hosts the implentation code, pre-trained weights, and dataset preparation details for the paper “TransWeather”. We also provide code for a strong transformer baseline for weather removal tasks.
Removing adverse weather conditions like rain, fog, and snow from images is an important problem in many applications. Most methods proposed in the literature have been designed to deal with just removing one type of degradation. Recently, a CNN-based method using neural architecture search (All-in-One) was proposed to remove all the weather conditions at once. However, it has a large number of parameters as it uses multiple encoders to cater to each weather removal task and still has scope for improvement in its performance. In this work, we focus on developing an efficient solution for the all adverse weather removal problem. To this end, we propose TransWeather, a transformer-based end-to-end model with just a single encoder and a decoder that can restore an image degraded by any weather condition. Specifically, we utilize a novel transformer encoder using intra-patch transformer blocks to enhance attention inside the patches to effectively remove smaller weather degradations. We also introduce a transformer decoder with learnable weather type embeddings to adjust to the weather degradation at hand. TransWeather achieves significant improvements across multiple test datasets over both All-in-One network as well as methods fine-tuned for specific tasks. In particular, TransWeather pushes the current state-of-the-art by +6.34 PSNR on the Test1 (rain+fog) dataset, +4.93 PSNR on the SnowTest100K-L dataset and +3.11 PSNR on the RainDrop test dataset. TransWeather is also validated on real world test images and found to be more effective than previous methods.
Using the code:
The code is stable while using Python 3.6.13, CUDA >=10.1
- Clone this repository:
git clone https://github.com/jeya-maria-jose/TransWeather cd TransWeather
To install all the dependencies using conda:
conda env create -f environment.yml conda activate transweather
If you prefer pip, install following versions:
timm==0.3.2 mmcv-full==1.2.7 torch==1.7.1 torchvision==0.8.2 opencv-python==126.96.36.199
TransWeather is trained on a combination of images sampled from Outdoor-Rain, Snow100K, and Raindrop datasets (similar to All-in-One (CVPR 2020)), dubbed as “All-Weather”, containing 18069 images. It can be downloaded from this link.
RainDrop Test : Link (Note that Test A is used for quantitative evaluation across all papers in the community, Test B is used for additional qualitative analysis)
Snow100K Test : Link (We use the Snow100K-L distribution for testing)
Test1 (validation set of “Outdoor-Rain”) : Link
Real World Images : Link
Download the datasets and arrange them in the following format. T
TransWeather ├── data | ├── train # Training | | ├── <dataset_name> | | | ├── input # rain images | | | └── gt # clean images | | └── dataset_filename.txt | └── test # Testing | | ├── <dataset_name> | | | ├── input # rain images | | | └── gt # clean images | | └── dataset_filename.txt
TransWeather Weights – Link
Place the folder in the root directory.
To run the evaluation for specific test datasets, run the following commands:
python test_snow100k.py -exp_name TransWeather_weights
python test_test1.py -exp_name TransWeather_weights
python test_raindropa.py -exp_name TransWeather_weights
These scripts will calculate the performance metrics as well as save the predictions in the results folder.
Training the network:
To train the network on All-weather dataset, run the following command:
python train.py -train_batch_size 32 -exp_name Transweather -epoch_start 0 -num_epochs 250
Note that Transweather is built to solve all adverse weather problem with a single model. We observe that, additionally TransWeather can be easilty modified (removing the transformer decoder) to just focus on a individual weather restoration task. To train just the Transweather-encoder on other datasets (like Rain-800), organize the dataset similar to all-weather and run the following command:
python train-individual.py -train_batch_size 32 -exp_name Transweather-finetune -epoch_start 0 -num_epochs 250
train-individual.py with the necesarry details of the data to be trained on. Note that the network used there is a sub-section of our original Transweather architecture without the weather queries.