SynthRSF: A Novel Photorealistic Synthetic Dataset for Adverse

Weather Condition Denoising

Angelos Kanlis

, Vazgken Vanian

, Sotiris Karvarsamis

, Ioanna Gkika

Konstantinos Konstantoudakis

and Dimitrios Zarpalas

Visual Computing Lab (VCL), Information Technologies Institute (ITI),

Centre for Research and Technology - Hellas (CERTH), Thessaloniki, Greece

Keywords:

Synthetic Dataset, Image Restoration, Adverse Weather Conditions, Semantic Segmentation, Depth

Estimation, Benchmarking, Unreal Engine.

Abstract:

This paper presents the SynthRSF dataset for training and evaluating single-image rain, snow and haze de-

noising algorithms, as well as evaluating object detection, semantic segmentation, and depth estimation per-

formance in noisy or denoised images. Our dataset features 26, 893 noisy images, each accompanied by its

corresponding ground truth image. It further includes 13, 800 noisy images accompanied by ground truth,

16-bit depth maps and pixel-accurate annotations for various object instances in each frame. The utility of

SynthRSF is assessed by training uniﬁed models for rain, snow, and haze removal, achieving good objective

metrics and excellent subjective results compared to existing adverse weather condition datasets. Further-

more, we demonstrate its use as a benchmark for the performance of an object detection algorithm in weather-

degraded image datasets.

1 INTRODUCTION

Creating scene understanding models has become a

central goal in both computer vision research and as-

sociated industrial applications. Such tasks can in-

volve object detection, segmentation, depth estima-

tion, as well as more complex procedures. However,

adverse conditions such as rain, snow and haze, as

well as variable lighting conditions, can impact the

performance of such algorithms by degrading the vi-

sual data. This can affect a wide range of applications,

such as autonomous driving, surveillance, robotics,

computer-assisted search-and-rescue, and more.

Due to the practical constraints of collecting

rain, snow and haze-speciﬁc data with an associated

ground truth at real-world sites, as well as the difﬁ-

culty of deﬁning the ground truth scene at a later time,

when lighting and other variables have changed, sig-

niﬁcant research effort has been devoted to generating

synthetic datasets for rain, snow and haze.

https://orcid.org/0009-0005-0836-9325

https://orcid.org/0000-0002-2150-3446

https://orcid.org/0000-0002-5302-3711

https://orcid.org/0000-0001-7340-3079

https://orcid.org/0000-0001-5092-8796

https://orcid.org/0000-0002-9649-9306

Although some datasets use a synthetic noise layer

superimposed on real world images, the result often

appears ﬂat and unconvincing to a human observer.

Furthermore, by simply layering weather noise on top

of an image, one does not account for the effect of

the weather phenomenon on the landscape nor the ef-

fect of existing lighting conditions on the appearance

of the weather phenomenon itself. For these reasons,

deep learning models trained on such datasets often

perform poorly in real-world conditions, as the do-

main gap between the training set and the actual input

in an application is signiﬁcant.

A solution to the above would be a photorealis-

tic synthetic dataset including adverse weather effects

as 3D effects fully integrated in a scene. In recent

years, modern game engines are capable of producing

highly realistic scenes, incorporating not only objects,

weather effects and lighting, but also their interaction.

Renders from such scenes can comprise multiple ver-

sions of the same view, including ones with adverse

weather conditions of various types and intensities, as

well as clear ground-truth images.

This paper presents SynthRSF (Synthetic with

Rain, Snow and Fog), a novel, photorealistic,

synthetic dataset focused on incorporating adverse

weather conditions, created using the Unreal 5.2 game

Kanlis, A., Vanian, V., Karvarsamis, S., Gkika, I., Konstantoudakis, K. and Zarpalas, D.

SynthRSF: A Novel Photorealistic Synthetic Dataset for Adverse Weather Condition Denoising.

DOI: 10.5220/0012397700003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 3: VISAPP, pages

567-574

ISBN: 978-989-758-679-8; ISSN: 2184-4321

567

engine. SynthRSF is based on 14 3D scenes, of var-

ious sizes (from indoor rooms to entire cities) set in

various environments (urban day, urban night, inte-

rior, nature), within which the camera moves on a

virtual rig, rendering images containing various types

of noise: snow, rain, uniform and non-uniform fog.

Each noisy frame is accompanied by the correspond-

ing ground truth image, for training denoising models.

To showcase SynthRSF’s added value in visual

computing research, a series of experiments has

been conducted, using it to train the state-of-the-

art TransWeather (Valanarasu et al., 2022) adverse

weather noise removal model. As a training dataset,

SynthRSF exhibits promising performance compared

to existing adverse weather datasets, and beyond

state-of-the-art performance when used in combina-

tion with some existing datasets. In addition, a human

subjective evaluation survey is performed, using real-

world images. Results provide compelling evidence

that when training models with photorealistic data,

denoising results are consistently deemed preferable

by human observers.

Furthermore, SynthRSF comes with an additional

multi-modal expansion dataset, named SynthRSF-

MM. The multi-modal dataset contains 14 scenes,

with pixel-level annotations for 5 object instances per

scene, and 41 object classes that are included in total.

With its additional modalities, it can be used as a train

and/or test dataset in a wider range of computer vi-

sion tasks, such as object detection, image segmenta-

tion, depth estimation, and scene understanding, with

possible applications in autonomous driving, robotics,

search-and-rescue, and more.

Hence, the main contributions of this paper can be

summarised as follows:

• the SynthRSF dataset, a synthetic photorealis-

tic dataset incorporating 3D weather effects and

lighting, comprising 26,893 pairs of images (de-

graded with adverse weather and ground truth).

SynthRSF, along with its expansion (see next bul-

let) is available on the Git repository

• the SynthRSF-MM expansion, an additional

13,800 pairs with ground truth on additional

modalities: depth map, semantic segmentation,

and bounding box pixel coordinates for 39 classes.

• a novel Dataset creation methodology based

on the Unreal 5.2 game engine, leveraging 3D

models and effects and predeﬁned virtual camera

paths, used to create SynthRSF/SynthRSF-MM.

• a set of Experiments comparing SynthRSF to

previously published adverse weather datasets in

training image restoration models.

https://github.com/VCL3D/SynthRSF

2 RELATED WORK

2.1 Uniﬁed Weather Denoising Models

Weather denoising research has recently undergone a

shift from earlier optimization-based techniques, of-

ten requiring priors that are tailored to speciﬁc types

of weather conditions, to deep-learning approaches

(Yang et al., 2019) can be used to model multiple phe-

nomena. The introduction of CNNs and GANs (Ren

et al., 2020) has signiﬁcantly enhanced denoising ca-

pabilities.

So far, the fraction of works on uniﬁed deraining,

desnowing and dehazing is still signiﬁcantly smaller

than the research work on rain, snow or haze in

isolation. However, very recently, multiple meth-

ods have emerged that follow a uniﬁed approach

(Valanarasu et al., 2022), (

Ozdenizci and Legenstein,

2022), (Wang et al., 2023), (Chen et al., 2022), (Kar-

avarsamis et al., 2022a). The authors of (Li et al.,

2020) are one of the ﬁrst to handle multiple weather

degradations using a single network. The model is

based on CNNs and consists of multiple weather-

speciﬁc encoders and a single common decoder.

(Valanarasu et al., 2022) proposes a single

encoder-decoder architecture based on transformers

and uses weather queries to handle multiple adverse

weather conditions. A novel transformer-based block

is also proposed improving the networks’ perfor-

mance. (Wang et al., 2023) is another work based

on transformers. To improve the learning capabili-

ties and efﬁciency of the model, transformers-based

blocks are used in a grid structure. The approach pro-

posed by (

Ozdenizci and Legenstein, 2022) is based

on denoising diffusion-based methods, introducing a

patch based diffusive restoration architecture enabling

arbitrary sized image processing.

2.2 Datasets Based on Real Images with

Synthetic Weather Effects

For each single phenomenon denoising task, most

methods use one or more of the following datasets,

which have been documented in (Yang et al., 2020)

and (Karavarsamis et al., 2022b).

For removing rain, datasets like Rain12600 (Fu

et al., 2017) and Rain12000 (Zhang and Patel, 2018)

have been widely used while similarly for snow,

Snow-100K (Liu et al., 2018) and CSD (Chen et al.,

2021) among others. When it comes to haze, notable

datasets include I-HAZE (Ancuti et al., 2018a) and

O-HAZE (Ancuti et al., 2018b). Recently, a novel

technique was published (Ba et al., 2022) for gener-

ating ground truth for rainy images. However, similar

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

568

approaches for other important weather conditions are

still missing in the scientiﬁc literature.

2.3 Fully Synthetic Game Engine

Datasets

There is a signiﬁcant amount of fully synthetic

datasets generated in game engines such as Blender

and Unity 3D. Important milestones include SceneNet

(A Handa, 2016) containing annotated 3D scenes

that can generate unlimited ground truth training

data, (Richter et al., 2016) who use game interac-

tion with graphics hardware to generate labeled data

and (Mayer et al., 2016) who provide a stereo video

dataset to estimate disparity and scene ﬂow. Fur-

thermore, (Butler et al., 2012) create an optical ﬂow

dataset derived from a 3D animated short ﬁlm. To our

knowledge, weather noise has not been implemented

in any fully synthetic game engine datasets.

3 THE SYNTHRSF DATASET

3.1 Design

The design goal of SynthRSF is to create a collec-

tion of photorealistic image pairs in different types of

environments, of weather-degraded images and their

corresponding clear ground-truth image.

This is achieved by adding 3D weather effects

simulating rain, snow and fog to realistic 3D Scenes.

This way, included weather noise can be parameter-

ized into numerous combinations resulting in a wide

range of visibility conditions.

Simulating real-life fog is particularly interesting,

since it is both a cause of occlusion and simultane-

ously it interacts with existing light sources, increas-

ing the illumination of parts of the scene. This type

of simulation,

2 3

has now been made possible with

state-of-the-art game engines.

3.2 Environment

The content environment of SynthRSF is based on 14

3D scenes designed in the Unreal 5.2 game engine

Scenes are sourced from the Unreal Engine’s doc-

umentation, including Unreal’s City Sample Project

and Hillside Project, which contribute most of the im-

ages of the dataset, due to the quality and variety of

the 3D assets contained. Other scenes are sourced

Lumen Global Illumination

Volumetric Fog

http://www.unrealengine.com/

from the documentation and the Unreal Marketplace,

all detailed on the SynthRSF Git repository

These scenes simulate a variety of environment

and lighting characteristics: urban/rural, day/night,

indoor/outdoor, wet/snowy, captured by a virtual

camera moving along set paths. Aiming for high vi-

sual realism, interior scenes do not include snow or

rain noise, but do include uniform or non-uniform

fog, as it approximates light smoke and can be use-

ful in emergency response applications.

Scenes 1–5 are divided into training (67%) and

test sets (33%). without risking data leakage, as the

camera does not revisit the same locations. Scenes

6-14 are entirely in the training set.

SynthRSF provides 26, 893 weather images of

rain, snow, uniform fog and non-uniform fog. This

novel addition of non-uniform fog is the reason for its

name including “fog” rather than “haze”. All images

are accompanied by their ground-truth pairs.

3.3 Weather Effect Implementation

Snow and rain are simulated in a virtual scene by

combining elements: Particles (sprites from Unreal’s

Niagara System) that represent close to medium-

distance occlusion and a fog component that sim-

ulates precipitation-induced light diffusion in larger

distances. Particle dimensions, velocity, angle, pop-

ulation and fog density are assigned sinusoidal func-

tions with different periods. Over enough time, all

their possible combinations are represented. Uniform

fog is created by Unreal Engine’s ExponentialHeight-

Fog module, while non-uniform fog was generated

using the Unreal’s Legacy particle system.

Blueprint functions for the snow and rain effects

as well as the custom rendering preset are included in

Git

3.4 SynthRSF-MM (Multi-Modal) Set

3.4.1 Motivation and Functionality

In 3D game engines, native data on each asset is avail-

able, hence additional ground-truth modalities can be

included in each sample.

SynthRSF-MM is an expansion to SynthRSF,

containing fewer samples but including additional

ground-truth modalities (depth, segmentation, and

object bounding boxes). It features 13, 800 noisy

images generated from 14 Unreal-Engine-sourced

scenes. SynthRSF-MM’s scenes have been manu-

ally populated with 39 classes of 3D objects, includ-

ing persons, vehicles, animals, etc. Subsequently,

3D rain, snow and fog effects are added to the

SynthRSF: A Novel Photorealistic Synthetic Dataset for Adverse Weather Condition Denoising

569

(a) Ground truth (b) Depth map (c) Semantic segmentation

(d) Fog (e) Rain (f) Snow

Figure 1: A sample scene from SynthRSF-MM. For each Ground truth image, there is one depth map, ﬁve pixel-level anno-

tations and 8 noisy images per phenomenon.

scene. Each image is accompanied by a 16-bit depth

map, pixel-accurate segmentation per object instance,

YOLO-compatible bounding box .json ﬁles.

Due to the manual labour involved, 825 unique

static camera views were set up, and 8 noisy images

per phenomenon (rain, snow, fog), per view, were

generated. Indoor scenes feature fog noise only.

Object detection bounding boxes are available for

calculating the accuracy of an object detection task.

SynthRSF-MM includes 39 of the YOLOv8

object

detector classes. Occluded objects with fewer than

100 pixels appearing on an image are not being allot-

ted a bounding box.

Having ground truth metadata, SynthRSF-MM

can be used in tasks besides image restoration, such as

semantic segmentation, object detection and distance

estimation, both in clear and degraded conditions.

4 EXPERIMENTS

In testing SynthRSF and comparing it to previous

datasets, TransWeather, a widely used state-of-the-art

deep learning model for weather noise removal is em-

ployed. Experiments compare the results of Trans-

Weather when trained by its default training dataset

(AllWeather), SynthRSF, or a combination of both.

Three different experiments were conducted:

1. Objective: PSNR and SSIM metrics.

2. Subjective comparison on real adverse weather

images.

https://github.com/ultralytics/ultralytics

3. Object detection: This experiment compares the

efﬁciency of YoloV8 on images restored by Trans-

Weather, trained on different datasets.

4.1 Training an Image Restoration

Network for Adverse Weather

Conditions

4.1.1 Architecture Selection

SynthRSF is suitable for uniﬁed bad weather removal

architectures (i.e. “all-in-one” models that can re-

move multiple weather conditions). The presence

of such publicly available models is rather limited.

Three publicly available uniﬁed models were iden-

tiﬁed and tested: TransWeather (Valanarasu et al.,

2022), WeatherDiffusion (

Ozdenizci and Legenstein,

2022), and AirNet (Li et al., 2022). Out of these,

TransWeather features good robustness combined

with a relatively fast training process; WeatherDiffu-

sion, although superior in quality, is extremely slow

both in training and inference, as also stated in the

publication itself; and AirNet proved to be unstable at

times due its contrastive learning approach.

Hence, the decision was made to employ Trans-

Weather alone to conduct our experiments, allow-

ing multiple training iterations and testing on tens of

thousands of samples.

4.1.2 Training and Testing Datasets

In the original publication, TransWeather is trained on

the AllWeather dataset, a combination of Snow100K

(Liu et al., 2018), Outdoor-Rain (Li et al., 2019) and

RainDrop (Qian et al., 2018) which contain images

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

570

Table 1: Quantitative results based on PSNR and SSIM for TransWeather trained on AllWeather, SynthRSF, their combination

and tested on Snow100K-L, test1 and SynthRSF test sets.

Train

Test

Snow100K-L

(AllWeather

snow)

test1

(AllWeather

rain and fog)

SynthRSF

Snow

SynthRSF

Rain

SynthRSF

Fog

Average

PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM

AllWeather 28.08 0.86 27.17 0.87 24.83 0.79 24.33 0.73 20.07 0.73 24.90 0.80

SynthRSF 19.85 0.69 15.33 0.60 27.68 0.85 27.89 0.83 25.17 0.82 23.18 0.76

AllWeather + SynthRSF 28.39 0.87 27.16 0.87 27.48 0.85 27.82 0.83 24.66 0.82 27.10 0.85

Noisy TransWeather

(AllWeather)

TransWeather

(SynthRSF)

TransWeather

(AllWeather+SynthRSF)

Figure 2: Qualitative results from denoising real world images: (a) Snow (Row 1), (b) Rain (Row 2), (c) Fog (Row 3).

with snow, fog/rain and raindrop degradations respec-

tively. In recent years AllWeather has become the go-

to dataset for uniﬁed models. As such, it was chosen

as the comparison dataset. For the evaluation pro-

cess three instances of TransWeather are utilised on

distinct datasets: (a) the original AllWeather dataset;

(b) the SynthRSF dataset; and, (c) the combination of

both datasets by using all images as input. Combin-

ing datasets often produces highly desirable results, as

the literature (Liu et al., 2019; Yao et al., 2023) sug-

gests. The three instances are evaluated on the combi-

nation of Snow-100KL (snow) and test1 (rain, fog), as

well as the test set of SynthRSF (snow, rain, fog). Im-

ages with raindrop noise were not used. For training,

SynthRSF images were downscaled from 1920x1080

to 720x405 pixels.

SynthRSF: A Novel Photorealistic Synthetic Dataset for Adverse Weather Condition Denoising

571

AllWeather Snow sample

AllWeather Rain-Fog sample

SynthRSF snow sample

SynthRSF rain sample

SynthRSF fog sample

Noisy Denoised

All-

Weather

Denoised

Synth-

RSF

Denoised

Both

Ground

truth

Figure 3: Denoising images from both datasets using

trained TransWeather models.

4.1.3 Training Parameters

While training, TransWeather the default settings

mentioned in (Valanarasu et al., 2022) are used with-

out alteration to any of the hyperparameters. All mod-

els are trained on a single NVIDIA RTX 3090 GPU

using the PyTorch framework (Paszke et al., 2019).

4.1.4 Quantitative Results

To evaluate performance we use the PSNR and SSIM

metrics. The results for Snow-100KL, test1 and

SynthRSF test set are summarized in Table 1. As

expected, the model instance trained on the combi-

nation of the datasets demonstrates the best overall

performance. Furthermore, using SynthRSF in com-

bination with AllWeather improves the performance

of TransWeather for Snow100KL while not hurting

the performance for test1. On the contrary, when

testing SynthRSF test set, AllWeather does not seem

to improve but rather diminishes the performance of

the model when used in combination with SynthRSF.

This demonstrates the efﬁcacy of SynthRSF when

combined with existing datasets and as well as its

comprehensiveness and efﬁciency as a standalone so-

lution.

4.1.5 Qualitative Results

Synthetic Datasets. The predictions of the three

model instances for images of the three test sets are il-

lustrated in Figure 3. Similarly to quantitative results,

the usage of SynthRSF appears to improve model per-

formance, especially for the case of fog removal.

Real World Images. In this case (Figure 2), the ef-

fectiveness of SynthRSF is more apparent. All model

instances showcase good denoising results, but the

instances that used SynthRSF either by itself or in

combination with AllWeather outperform the one that

uses AllWeather only. Notably, the instance trained

solely on SynthRSF removes even the farther and

denser fog element, while the instance trained solely

on AllWeather struggles to remove fog even at close

distances.

4.2 Subjective Assessment Experiment

In order to evaluate training by SynthRSF in compar-

ison with previous datasets, tests are performed on 75

real-world images collected from the Internet. Al-

though a ground truth clear image for such cannot

exist, and hence numerical results are not applicable,

qualitative comparison can be performed on a subjec-

tive level.

Those images were fed to the three previously

trained TransWeather models, and the results were

evaluated by 70 survey participants. Participants, us-

ing their personal displays, with no time restrictions,

accessed an online form to compare each noisy image

with the three randomly ordered denoised versions,

selecting the one they found clearest.

Survey results in Table 2 show a strong preference

for SynthRSF-restored images, especially in images

of rain and fog. While preferences for ”snowy” im-

ages are less pronounced, SynthRSF still leads. De-

spite the model trained on AllWeather often removing

more individual snowﬂakes or rain streaks, the model

trained on SynthRSF, because of its fog data, tends

to clean up the distant parts of the image that are ob-

scured by the fog-induced light scattering.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

572

Table 2: Subjective assessment experiment - Summary of

data from 70 survey participants for the total votes per

model, and the number of times each model was chosen

as the preferred option across 75 noisy images.

Training set All-

Weather

Synth-

RSF

Both

Rain

Total votes 149/1167 610/1167 408/1167

Top voted image 1/21 12/21 8/21

Snow

Total votes 381/1773 795/1773 597/1773

Top voted image 3/30 15/30 12/30

Fog

Total votes 149/1567 911/1567 507/1567

Top voted image 1/24 18/24 5/24

4.3 Benchmarking an Object Detector

on Denoised Images

4.3.1 Benchmarking YoloV8 on Synthetic Noisy

Images

As a test case for SynthRSF-MM, it is used to bench-

mark the performance of YOLOv8 on images con-

taining adverse weather noise and their denoised

counterparts. As testing data, the COCO validation

dataset is used, with overlayed snow masks from

CSD (Chen et al., 2021) and SRRS (Chen et al.,

2020) datasets, as well as rain masks from RainTrainL

(Zhang and Patel, 2018) dataset. For fog, the COCO

dataset does not provide depth maps, so the RTTS

dataset is used (Li et al., 2018) providing noisy im-

ages and object annotations. Results are summarised

in Table 3.

4.3.2 Benchmarking YoloV8 on SynthRSF-MM

Demonstrating the utility of SynthRSF-MM’s annota-

tions, the denoising and object detection experiment

is performed on the SynthRSF-MM images. The re-

sults are summarised in table 4. The combined All-

Weather+SynthRSF dataset training produces better

results in snow and rain images, while SynthRSF-

only training was the most beneﬁcial in fog images.

5 CONCLUSIONS

In this paper, we have presented and shared

SynthRSF, a novel synthetic dataset focused on ad-

verse conditions. Its utility has been validated in mul-

tiple experiments: a) by training the TransWeather

Table 3: YOLOV8 results on COCO/RTTS with syn-

thetic noise, denoised by TransWeather trained on different

datasets. Combined training is superior in snow and rain,

SynthRSF performs better alone in fog.

Training set mAP50 mAP50-95

Snow (COCO w/CSD/SRRS)

noisy 0.623 0.466

AllWeather 0.629 0.468

SynthRSF 0.605 0.447

AllWeather+SynthRSF 0.632 0.471

Rain (COCO w/RainTrainL)

noisy 0.626 0.466

AllWeather 0.615 0.455

SynthRSF 0.616 0.457

AllWeather+SynthRSF 0.628 0.466

Fog (RTTS)

noisy 0.656 0.416

AllWeather 0.644 0.409

SynthRSF 0.665 0.42

AllWeather+SynthRSF 0.658 0.417

Table 4: YOLOV8 results on SynthRSF-MM dataset de-

noised by TransWeather trained on different datasets.

Training set mAP50 mAP50-95

Rain

noisy 0.293 0.218

AllWeather 0.319 0.239

SynthRSF 0.327 0.25

AllWeather+SynthRSF 0.336 0.256

Snow

noisy 0.303 0.224

AllWeather 0.326 0.245

SynthRSF 0.314 0.238

AllWeather+SynthRSF 0.325 0.247

Fog

noisy 0.296 0.22

AllWeather 0.288 0.217

SynthRSF 0.307 0.227

AllWeather+SynthRSF 0.302 0.227

image denoising deep learning model in a series of

both objective and subjective experiments; and b) by

benchmarking a state-of-the-art object detection algo-

rithm in its performance in the absence or presence of

adverse weather conditions.

We have also presented SynthRSF-MM, a novel

multi-modal dataset, which includes depth maps for

all images, as well as pixel-level annotations for 39

object classes. Although its potential uses are many,

the experiments highlight its functionality as a test set

for measuring an object detector’s performance in var-

ious inputs.

SynthRSF: A Novel Photorealistic Synthetic Dataset for Adverse Weather Condition Denoising

573

ACKNOWLEDGEMENTS

This research has been supported by the Euro-

pean Commission funded program RESCUER, under

H2020 Grant Agreement 101021836.

REFERENCES

A Handa, V Patraucean, V. B. S. S. R. C. (2016). Under-

standing real world indoor scenes with synthetic data.

In CVPR 2016.

Ancuti, C., Ancuti, C. O., Timofte, R., and

De Vleeschouwer, C. (2018a). I-haze: a dehaz-

ing benchmark with real hazy and haze-free indoor

images. In ACIVS 2018, Proceedings 19. Springer.

Ancuti, C. O., Ancuti, C., Timofte, R., and

De Vleeschouwer, C. (2018b). O-haze: A de-

hazing benchmark with real hazy and haze-free

outdoor images. In CVPR 2018, NTIRE Workshop,

NTIRE CVPR’18, Salt Lake City, Utah, USA.

Ba, Y., Zhang, H., Yang, E., Suzuki, A., Pfahnl, A., Chan-

drappa, C. C., de Melo, C. M., You, S., Soatto, S.,

Wong, A., et al. (2022). Not just streaks: Towards

ground truth for single image deraining. In ECCV.

Butler, D. J., Wulff, J., Stanley, G. B., and Black, M. J.

(2012). A naturalistic open source movie for optical

ﬂow evaluation. In ECCV Proceedings, Part VI 12.

Chen, W.-T., Fang, H.-Y., Ding, J.-J., Tsai, C.-C., and Kuo,

S.-Y. (2020). Jstasr: Joint size and transparency-

aware snow removal algorithm based on modiﬁed par-

tial convolution and veiling effect removal. In ECCV

2020 Proceedings, Part XXI 16. Springer.

Chen, W.-T., Fang, H.-Y., Hsieh, C.-L., Tsai, C.-C., Chen,

I., Ding, J.-J., Kuo, S.-Y., et al. (2021). All snow re-

moved: Single image desnowing algorithm using hier-

archical dual-tree complex wavelet representation and

contradict channel loss. In CVPR 2021.

Chen, W.-T., Huang, Z.-K., Tsai, C.-C., Yang, H.-H., Ding,

J.-J., and Kuo, S.-Y. (2022). Learning multiple ad-

verse weather removal via two-stage knowledge learn-

ing and multi-contrastive regularization: Toward a

uniﬁed model. In CVPR 2022.

Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., and Pais-

ley, J. (2017). Removing rain from single images via

a deep detail network. In CVPR 2017.

Karavarsamis, S., Doumanoglou, A., Konstantoudakis, K.,

and Zarpalas, D. (2022a). Cross-stitched multi-task

dual recursive networks for uniﬁed single image de-

raining and desnowing. In WF-IoT 2022.

Karavarsamis, S., Gkika, I., Gkitsas, V., Konstantoudakis,

K., and Zarpalas, D. (2022b). A survey of deep

learning-based image restoration methods for enhanc-

ing situational awareness at disaster sites: the cases of

rain, snow and haze. Sensors, 22(13).

Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., and Peng, X. (2022).

All-in-one image restoration for unknown corruption.

In CVPR 2022.

Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., and

Wang, Z. (2018). Benchmarking single-image dehaz-

ing and beyond. IEEE Transactions on Image Pro-

cessing, 28(1).

Li, R., Cheong, L.-F., and Tan, R. T. (2019). Heavy rain

image restoration: Integrating physics model and con-

ditional adversarial learning. In CVPR 2019.

Li, R., Tan, R. T., and Cheong, L.-F. (2020). All in one bad

weather removal using architectural search. In CVPR.

Liu, P., Zhou, X., Yang, J., El Basha Mohammad, D., and

Fang, R. (2019). Image restoration using deep regu-

lated convolutional networks.

Liu, Y.-F., Jaw, D.-W., Huang, S.-C., and Hwang, J.-N.

(2018). Desnownet: Context-aware deep network for

snow removal. IEEE Trans. Image Processing, 27(6).

Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D.,

Dosovitskiy, A., and Brox, T. (2016). A large dataset

to train convolutional networks for disparity, optical

ﬂow, and scene ﬂow estimation. In CVPR 2016.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., Desmaison, A., Kopf, A., Yang, E., De-

Vito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,

Steiner, B., Fang, L., Bai, J., and Chintala, S. (2019).

PyTorch: An Imperative Style, High-Performance

Deep Learning Library. In Advances in Neural Infor-

mation Processing Systems 32.

Qian, R., Tan, R. T., Yang, W., Su, J., and Liu, J. (2018).

Attentive generative adversarial network for raindrop

removal from a single image. In CVPR 2018.

Ren, Y., Li, S., Nie, M., and Chuankun, L. (2020). Single

image de-raining via improved generative adversarial

nets. Sensors, 20.

Richter, S. R., Vineet, V., Roth, S., and Koltun, V. (2016).

Playing for data: Ground truth from computer games.

In ECCV 2016 Proceedings, Part II 14. Springer.

Valanarasu, J. M. J., Yasarla, R., and Patel, V. M. (2022).

Transweather: Transformer-based restoration of im-

ages degraded by adverse weather conditions. In

CVPR 2022.

Wang, T., Zhang, K., Shao, Z., Luo, W., Stenger, B., Lu, T.,

Kim, T.-K., Liu, W., and Li, H. (2023). Gridformer:

Residual dense transformer with grid structure for im-

age restoration in adverse weather conditions. arXiv

preprint arXiv:2305.17863.

Yang, W., Tan, R. T., Wang, S., Fang, Y., and Liu, J. (2019).

Single image deraining: From model-based to data-

driven and beyond.

Yang, W., Tan, R. T., Wang, S., Fang, Y., and Liu, J. (2020).

Single image deraining: From model-based to data-

driven and beyond. PAMI 2020, 43(11).

Yao, M., Xu, R., Guan, Y., Huang, J., and Xiong, Z. (2023).

Neural degradation representation learning for all-in-

one image restoration.

Zhang, H. and Patel, V. M. (2018). Density-aware single

image de-raining using a multi-stream dense network.

In CVPR 2018.

Ozdenizci, O. and Legenstein, R. (2022). Restoring vi-

sion in adverse weather conditions with patch-based

denoising diffusion models.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

574