Adherent raindrops on camera lenses are a pervasive source of real-world image degradation, undermining outdoor visual perception for downstream applications such as robotics and 3D scene reconstruction. However, most existing raindrop datasets and methods target only background-focused 2D images, overlooking the equally common raindrop-focused case and the impact of raindrops on 3D scene reconstruction. Moreover, they are almost exclusively daytime, lacking the complex artificial lighting and reflections of nighttime raindrop scenes. To address these issues, we introduce OmniRaindrop, a real-world dataset and benchmark that jointly evaluates 2D raindrop removal and 3D scene reconstruction across raindrop-focused and background-focused captures, under both daytime and nighttime conditions. OmniRaindrop provides spatially aligned raindrop-focused, background-focused, and rain-free ground-truth image sets: a 2D subset of 15,186 aligned pairs/triplets and a 3D subset of 40 real-world scenes evenly split between day and night, totaling 3,822 multi-view images. On top of this data, the benchmark traces how raindrop degradation propagates from 2D restoration to 3D reconstruction by testing representative restoration methods and their effects on pose estimation, point cloud initialization, and 3DGS optimization. Comprehensive experiments reveal that raindrop-focused scenes are substantially harder than background-focused ones, and that errors in 2D restoration and pose/point cloud initialization cascade into degraded 3D reconstruction, pointing to clear directions for more robust 2D restoration and 3DGS design. Building on the dataset, we have organized two consecutive NTIRE challenges (2025 and 2026) on Day and Night Raindrop Removal for Dual-Focused Images.
Figures & tables
Figure 1: Motivation for OmniRaindrop. Top block: background-focused captures (blurry raindrops) Qian et al. (2018) ; Soboleva and Shipitko (2021) ; Jin et al. (2024a) ; bottom block: raindrop-focused captures (blurry background); within each block, the upper and lower rows show daytime and nighttime, respectively Jin et al. (2024a) .
Figure 2: Illustration of the OmniRaindrop dataset. OmniRaindrop covers day/night and raindrop-/background-focused settings, providing aligned 2D pairs/triplets and real 3D multi-view scenes with rain-free references for evaluating 2D raindrop removal and 3DGS reconstruction.
Figure 3: Physical formulation of raindrop-focused degradation. Given a raindrop-focused image x~ and its corresponding rain-free blurred background x0 , the difference map m~ is computed to isolate raindrop-induced degradation, while b0 denotes the sharp rain-free ground truth.
Figure 4: From 2D to 3D raindrop removal benchmark pipeline. We develop the first benchmark for comprehensively evaluating raindrop removal performance. The benchmark begins with dataset collection, proceeds through data processing, and ends with a raindrop-aware 3DGS evaluation. Our benchmark enables the community to explore background-focused and raindrop-focused images, including challenges unique to daytime and nighttime conditions.
Dataset
Components
Scale
Source
Bg-focus
Rd-focus
Scenes
Images
Day
Night
Day
Night
2D Raindrop Datasets
Qian et al. Qian et al. (2018)
919 Pairs
−
1,838
Real
✓
×
×
×
RainDS-Real Quan et al. (2021)
248 Quadruplets
−
992
Real
✓
×
×
×
RainDS-Syn Quan et al. (2021)
1,200 Quadruplets
−
4,800
Syn
✓
×
×
×
RobotCar Porav et al. (2019)
4,818 Pairs
−
9,636
Real
✓
×
×
×
Table 1: Comprehensive comparison between OmniRaindrop and existing datasets. We compare dataset components, scale, data source, focus condition, and illumination coverage. OmniRaindrop provides a large-scale real-world benchmark with strict alignment for both 2D restoration and 3DGS reconstruction.
Data/Metric
Input
Raindrop Algorithms
Restoration Backbones
AtGAN
Robot
RainDS
IDT
RDiff
Restormer
Uformer
DiT
Day
PSNR ↑
21.92
23.62
23.82
25.39
26.05
25.52
26.08
25.78
26.03
SSIM ↑
0.560
0.658
0.670
0.717
0.736
0.734
0.748
0.727
0.752
LPIPS ↓
0.247
0.200
0.191
0.147
0.141
0.111
0.131
0.148
0.106
Night
PSNR ↑
24.78
24.38
24.57
25.39
26.81
26.48
26.87
25.26
26.23
SSIM ↑
0.726
0.773
0.787
0.820
0.851
0.831
0.848
0.816
0.826
Table 2: Quantitative evaluation of representative raindrop removal methods and general image restoration backbones on the daytime and nighttime subsets of OmniRaindrop. We report PSNR, SSIM, and LPIPS to evaluate pixel fidelity, structural similarity, and perceptual quality, respectively. The best results are highlighted for each metric.
Figure 5: Qualitative comparison of representative raindrop removal methods on daytime images. Existing methods remove some diffuse occlusions in background-focused inputs, but struggle with the sharper raindrops and severe background blur in raindrop-focused inputs.
Figure 6: Qualitative comparison of representative raindrop removal methods on nighttime images. Compared with daytime scenes, nighttime raindrops introduce stronger light refraction, specular highlights, halos, local saturation, and non-linear intensity variations. Existing methods often leave residual raindrops or over-smooth background details in background-focused images. Under raindrop-focused inputs, the combined effect of sharp raindrop structures, severe background defocus, and nighttime illumination distortion makes it more difficult to recover clean and structurally faithful backgrounds.
3DGS
3DGS-MCMC
GS-W
WeatherGS
Scene
Metrics
Uformer
Restormer
IDT
Uformer
Restormer
IDT
Uformer
Restormer
IDT
/
Day
PSNR ↑
17.119
17.162
17.211
18.202
18.217
18.217
19.005
18.983
19.192
16.918
SSIM ↑
0.468
0.470
0.476
0.501
0.494
0.502
0.540
0.538
0.550
0.443
LPIPS ↓
0.442
0.439
0.434
0.419
0.419
0.414
0.355
0.355
0.340
0.446
Night
PSNR ↑
20.633
20.744
20.862
21.222
21.476
21.537
22.254
22.268
22.162
20.672
SSIM ↑
0.536
0.540
0.544
0.622
0.620
0.622
0.596
0.588
0.580
0.518
Table 3: Quantitative comparison of four 3DGS methods on the 3D subset using VGGT Wang et al. (2025a) for camera pose estimation and point cloud initialization. For each daytime/nighttime split, results are averaged over all background-focused and raindrop-focused cases. Uformer Wang et al. (2022) , Restormer Zamir et al. (2022) , and IDT Xiao et al. (2022) denote different 2D restoration front-ends used before 3DGS reconstruction, while WeatherGS Qian et al. (2025) is evaluated with its original pipeline.
Figure 7: Qualitative comparison of four 3DGS methods under raindrop-corrupted inputs. Background-focused inputs generally preserve more scene details and lead to higher reconstruction quality than raindrop-focused inputs. WeatherGS Qian et al. (2025) produces relatively smooth reconstructions for blurred inputs, while GS-W Zhang et al. (2024) produces the most visually coherent results in these examples.
Figure 8: Qualitative comparison of four 3DGS methods under nighttime raindrop-corrupted inputs. Nighttime scenes introduce complex artificial-light refraction, highlights, and low-light degradation, which can introduce inconsistencies across views. Compared with background-focused inputs, raindrop-focused inputs lead to more severe structural degradation due to sharp raindrop interference and defocused background observations.
Method
background-
raindrop-
focus
focus
Empty Scene
Avg. Points
Empty Scene
Avg. Points
COLMAP
0/40
6,199.55
1/40
1,634.54
VGGT
9/40
82,069.97
37/40
6,402.00
Table 4: Comparison of point cloud initialization generated by COLMAP Schonberger and Frahm (2016) and VGGT Wang et al. (2025a) under background-focus and raindrop-focus inputs. Empty scenes denote completed runs that produce zero point or invalid initial point cloud files. Average point counts are computed only over scenes with valid point clouds.
Figure 9: Camera pose quality comparison between COLMAP Schonberger and Frahm (2016) and VGGT Wang et al. (2025a) under background-focused and raindrop-focused inputs. We report four metrics across representative daytime and nighttime scenes: (a) registration rate, (b) rotation error, (c) translation error, and (d) AUC@30. Raindrop-focused inputs substantially reduce registration stability and pose accuracy, especially for COLMAP, while VGGT shows stronger robustness but still degrades under severe raindrop-focused conditions.
Input Image
Geometry Initialization
PSNR ↑
SSIM ↑
LPIPS ↓
Raindrop
Raindrop
19.037
0.510
0.415
Raindrop
GT-image
21.310
0.620
0.363
GT
Raindrop
20.464
0.551
0.333
GT
GT-image
25.525
0.811
0.156
Table 5: Ablation study of appearance inputs and geometry initialization for 3DGS reconstruction. In the Input Image column, “Raindrop” denotes raindrop-corrupted images restored by IDT Xiao et al. (2022) before being used for 3DGS optimization, whereas “GT” denotes the corresponding rain-free ground-truth images. In the Geometry Initialization column, “Raindrop” and “GT” indicate that camera poses and initial point clouds are estimated from the original, unprocessed raindrop-corrupted images and the rain-free ground-truth images, respectively.
This paper presents an overview of the NTIRE 2026 Second Challenge on Day and Night Raindrop Removal for Dual-Focused Images. Building upon the success of the first edition, this challenge attracted a wide range of impressive solutions, all developed and evaluated on our real-world Raindrop Clarity dataset~\cite{jin2024raindrop}. For this edition, we adjust the dataset with 14,139 images for training, 407 images for validation, and 593 images for testing. The primary goal of this challenge is to establish a strong and practical benchmark for the removal of raindrops under various illumination and focus conditions. In total, 168 teams have registered for the competition, and 17 teams submitted valid final solutions and fact sheets for the testing phase. The submitted methods achieved strong performance on the Raindrop Clarity dataset, demonstrating the growing progress in this challenging task.
When capturing images through glass surfaces or windshields on rainy days, raindrops and reflections frequently co-occur to significantly reduce the visibility of captured images. This practical problem lacks attention and needs to be resolved urgently. Prior de-raindrop, de-reflection, and all-in-one models have failed to address this composite degradation. To this end, we first formally define the unified removal of raindrops and reflections (UR3) task for the first time and construct a real-shot dataset, namely RainDrop and ReFlection (RDRF), which provides a new benchmark with substantial, high-quality, diverse image pairs. Then, we propose a novel diffusion-based framework (i.e., DiffUR3) with several target designs to address this challenging task. By leveraging the powerful generative prior, DiffUR3 successfully removes both types of degradations. Extensive experiments demonstrate that our method achieves state-of-the-art performance on our benchmark and on challenging in-the-wild images.
Xingyu Liu, Zewei He, Yu Chen +4
Huanjiang Laboratory, Zhuji, China · School of Aeronautics and Astronautics, Zhejiang University, Hangzhou, China · Hangzhou Institute of Technology, Xidian University, Hangzhou, China +1
Raindrops adhered to camera lens or windshield are inevitable in rainy scenes and can become an issue for many computer vision systems such as autonomous driving. Because raindrop appearance is affected by too many parameters, therefore it is unlikely to find an effective model based solution. Learning based methods are also problematic, because traditional learning method cannot properly model the complex appearance. Whereas deep learning method lacks sufficiently large and realistic training data. To solve it, in our work, we propose the first photo-realistic dataset of synthetic adherent raindrops for training. The rendering is physics based with consideration of the water dynamic, geometric and photometry. The dataset contains various types of rainy scenes and particularly the rainy driving scenes. Based on the modeling of raindrop imagery, we introduce a detection network which has the awareness of the raindrop refraction as well as its blurring. Based on that, we propose the removal network that can well recover the image structure. Rigorous experiments demonstrate the state-of-the-art performance of our proposed framework.
Zhixiang Hao, Shaodi You, Yu Li +2
State Key Laboratory of VR Technology and Systems, Beihang University, Beijing, China · 61-CSIRO · Tencent +2