SynthRSF Dataset

Paper

SynthRSF: A Novel Photorealistic Synthetic Dataset for AdverseWeather Condition Denoising

Overview

• SynthRSF (Synthetic with Rain, Snow, uniform and non-uniform Fog) dataset is introduced for training and evaluating adverse weather image denoising models as well as use in object detection, semantic segmentation, and depth estimation models.
• SynthRSF addresses a gap in synthetic datasets for adverse weather conditions, contributing significantly more photorealistic data compared to common 2D layered noise datasets, as well as additional modalities.
• Applications include autonomous driving, surveillance, robotics, computer-assisted search-and-rescue.

Contents

• SynthRSF:

  • 26,893 photorealistic image pairs (noisy and ground truth).
  • 14 3D scenes set in various environmental (rural/urban), contextual (indoor/outdoor) and lighting conditions (day/night).
  • Created using Unreal 5.2 engine.

• SynthRSF-MM expansion:

  • 13,800 additional pairs are accompanied by:
  • 16-bit depth maps.
  • Pixel-accurate object annotations for 41 object classes.

SynthRSF Examples

Rain

Snow

Uniform Fog

Non-Uniform Fog

SynthRSF-MM Dataset

Overview

SynthRSF-MM is an expansion of the SynthRSF dataset, including additional modalities for a comprehensive analysis in computer vision tasks. It contains 13,800 noisy images from 14 scenes with ground truth on additional modalities.

SynthRSF-MM Example

Ground truth	Depth map	Object annotations
Fog	Rain	Snow

A sample scene from SynthRSF-MM, showing the ground truth (clear) scene, depth map, semantic segmentation stencil mask, and three noisy variants with different weather conditions. For each Ground truth image, there is one depth map, five pixel-level annotations and 8 noisy images per phenomenon.

---

SynthRSF Source 3D Scenes

No	Description	Snow	Rain	Fog (U)	Fog (NU)
1	CitySample day	3862	10758	522	430
2	CitySample night	2695	2744	-	-
3	Hillside Sample	485	412	254	308
4	AngryMesh W1	279	287	180	211
5	AngryMesh W2	400	377	171	170
6	ICVFXProdTest	-	-	8	14
7	Realistic Rend.	-	-	75	75
8	Archviz Interior	-	-	-	30
9	Factory	-	-	13	16
10	Office Vol. 1	-	-	45	38
11	Bedroom Vol. 1	-	-	246	203
12	Kitchenette	-	-	3	15
13	Megascans A. A.	-	-	19	42
14	MeerkatDemo	250	250	51	55
	Total	7971	14828	1487	1607

---

SynthRSF-MM Source 3D Scenes

No	Description	Ground Truth	Depth	Snow	Rain	Fog (NU)
1	CitySample day	226	226	1808	1808	1808
2	CitySample night	158	158	1264	1264	1264
3	Archviz Interior	32	32	-	-	256
4	Factory	32	32	-	-	256
5	OfficeVol1	102	102	-	-	816
6	Bedroom Vol 1	22	22	-	-	176
7	Kitchenette	22	22	-	-	176
8	MegascansAA	20	20	-	-	160
9	MegascansSnow	20	20	160	160	160
10	Desert	17	17	136	136	136
11	MeerkatDemo	21	21	168	168	168
12	ICVFXProdTest	21	21	168	168	168
13	AngryMesh W1	21	21	168	168	168
14	AngryMesh W2	21	21	168	168	168
	Total	825	825	4040	4040	5720

---

Access

Dataset published under CC license (CC BY 4.0), free to use for all purposes, and requiring attribution: The dataset is available on Google Drive: https://drive.google.com/drive/folders/1UiXTYSIWbX2w62E-WDjxMyDGWvWDF0oX?usp=drive_link

Paper published under CC license (CC BY-NC-ND 4.0), free to use for non-commercial purposes, and requiring attribution: SynthRSF: A Novel Photorealistic Synthetic Dataset for AdverseWeather Condition Denoising

Kanlis, A.; Vanian, V.; Karvarsamis, S.; Gkika, I.; Konstantoudakis, K. and Zarpalas, D. (2024). SynthRSF: A Novel Photorealistic Synthetic Dataset for Adverse Weather Condition Denoising. In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 3: VISAPP, ISBN 978-989-758-679-8, ISSN 2184-4321, pages 567-574. DOI: 10.5220/0012397700003660

---