Awesome
Syn2Real
Syn2Real Transfer Learning for Image Deraining using Gaussian Processes
Rajeev Yasarla*, Vishwanath A. Sindagi*, Vishal M. Patel
Paper Link(CVPR '20)
@InProceedings{Yasarla_2020_CVPR,
author = {Yasarla, Rajeev and Sindagi, Vishwanath A. and Patel, Vishal M.},
title = {Syn2Real Transfer Learning for Image Deraining Using Gaussian Processes},
booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}
We propose a Gaussian Process-based semi-supervised learning framework which enables the network in learning to derain using synthetic dataset while generalizing better using unlabeled real-world images. Through extensive experiments and ablations on several challenging datasets (such as Rain800, Rain200H and DDN-SIRR), we show that the proposed method, when trained on limited labeled data, achieves on-par performance with fully-labeled training. Additionally, we demonstrate that using unlabeled real-world images in the proposed GP-based framework results in superior performance as compared to existing methods.
Journal extension:
Semi-Supervised Image Deraining using Gaussian Processes
Prerequisites:
- Linux
- Python 2 or 3
- Pytorch version >=1.9
- CPU or NVIDIA GPU + CUDA CuDNN (CUDA 10.2)
Dataset structure
- download the rain datasets and arrange the rainy images and clean images in the following order
- Save the image names into text file (dataset_filename.txt)
.
├── data
| ├── train # Training
| | ├── derain
| | | ├── <dataset_name>
| | | | ├── rain # rain images
| | | | └── norain # clean images
| | | └── dataset_filename.txt
| └── test # Testing
| | ├── derain
| | | ├── <dataset_name>
| | | | ├── rain # rain images
| | | | └── norain # clean images
| | | └── dataset_filename.txt
To test Syn2Real:
- mention test dataset text file in the line 57 of test.py, for example
val_filename = 'SIRR_test.txt'
- Run the following command
python test.py -category derain -exp_name DDN_SIRR_withGP
To train Syn2Real:
- mention the labeled, unlabeled, and validation dataset in lines 119-121 of train.py, for example
labeled_name = 'DDN_100_split1.txt'
unlabeled_name = 'real_input_split1.txt'
val_filename = 'SIRR_test.txt'
- Run the following command to train the base network without Gaussian processes
python train.py -train_batch_size 2 -category derain -exp_name DDN_SIRR_withoutGP -lambda_GP 0.00 -epoch_start 0
- Run the following command to train Syn2Real (CVPR'20) model
python train.py -train_batch_size 2 -category derain -exp_name DDN_SIRR_withGP -lambda_GP 0.0015 -epoch_start 0 -version version1
- Run the following command to train Syn2Real++ (journal submission GP modellig at feature map level)
python train.py -train_batch_size 2 -category derain -exp_name DDN_SIRR_withGP -lambda_GP 0.0015 -epoch_start 0 -version version2
Cross-domain experiments and Gaussian kernels
cross domain experiments are performed using DIDMDN dataset as source dataset, and other datasets like Rain800, JORDER_200L, DDN as target datasets.
----------------------------------------------------
Source datasets | Target datasets |
----------------------------------------------------
DIDMDN | Rain800, JORDER_200L, DDN |
----------------------------------------------------
Gaussian processes can be modelled using different kernels like Linear or Squared_exponential or Rational_quadratic. the updated code provides an option to choose the kernel type
-kernel_type <Linear or Squared_exponential or Rational_quadratic>
Fast version of GP
use GP_new_fast.py file for faster version of GP.
To use this GP_new_fast.py :
comment line 14 in train.py
and uncomment line 15 in train.py
Additionally you can use "train_new_comb.py" instead of "train.py".
In "train_new_comb.py" does iterative training of the network, i.e. each iteration contains one labeled train step and one unlabeled train step.
Run the following command to train Syn2Real (CVPR'20) model using "train_new_comb.py".
python train_new_comb.py -train_batch_size 2 -category derain -exp_name DDN_SIRR_withGP -lambda_GP 0.0015 -epoch_start 0 -version version1