Source-linked AI summary
LoViF 2026 The First Challenge on Unified Removal of Raindrops and Reflections: Methods and Results
Zewei He, Xi Tong, Yu Chen, Xingyu Liu, Xin Li, Zepeng Wang, Jiagao Hu, Fuhao Li, Yuxuan Chen, Fei Wang, Daiguo Zhou, Minmin Yi, Chuanrui Zhang, Liwen Zhang, Yeongjin Jeong, Hyunjin Cho, Jiwon Lee, Minsang Kim, Jae Woong Soh, Jin-Hui Jiang, Rong-Lin Jian, Chih-Chung Hsu, Youngjin Oh, Junhyeong Kwon, Junyoung Park, Jae Hyun Park, Sung Ju Lee, Nam Ik Cho, Vishwajeet Shukla, Himanshu Baurai, Zhiqi Zhang, Kui Jiang, Zhaocheng Yu, Runzhe Li, Dawei Fan, Hao Li, Zhanshuo Zhang, Fan Ji, Jiangmeng Li, Xiongxin Tang, Fanjiang Xu, Shangquan Sun, Anh-Kiet Duong, Petra Gomez-Krämer, Jean-Michel Carozza, Ruibo Zhang, Dexiang Hong, Xinyan Liu, Shengeng Tang, Weidong Chen, Tzu-Hsuan Weng, Min-Te Sun
TL;DR
The paper examines unified removal of raindrop-reflection composite degradation, a practical problem for rainy autonomous-driving imagery that prior separate-task methods often fail to address simultaneously. It reviews the first UR3 Challenge, its RDRF-based evaluation, submitted methods, and results. Xedit Master achieved the best overall challenge score, while the authors identify multi-frame fusion as inconsistent with the intended single-image setting.
Problem
Raindrops and reflections frequently co-occur in rainy autonomous-driving scenes, while methods designed for separate degradations often fail to remove both simultaneously.
Method
The paper reviews the first UR3 Challenge, using the real-shot RDRF dataset and analyzing submitted unified-restoration methods and challenge results.
Results
Xedit Master achieved the best overall challenge performance with a final score of 41.8226 and the highest PSNR of 32.3993.
Takeaways & Limitations
The challenge provides a benchmark for unified raindrop and reflection removal and highlights effective approaches for this composite degradation.
Takeaways & Limitations
Multi-frame fusion is inconsistent with the intended single-image task because the dataset acquisition protocol maps one ground-truth image to multiple degraded images.
Abstract
from arXiv · showhide
This workshop paper comprehensively reviews the First Challenge on Unified Removal of Raindrops and Reflections. The challenge aims to address a frequently encountered practical problem in the field of autonomous driving, i.e., raindrop-reflection composite degradation on rainy days. This competition attracted 149 registered participants and received 12 valid final submissions with corresponding fact sheets, significantly contributing to the progress of unified removal of raindrops and reflections. All the methods are developed and evaluated on our real-shot RainDrop and ReFlection (RDRF) dataset. A detailed analysis of the submitted methods and corresponding results is provided in this report, which highlights effective approaches and provides interesting insights for future research.
1 Introduction
Rainy autonomous-driving scenes often combine raindrops and camera-side reflections, reducing image visibility and potentially creating safety hazards. The challenge targets unified removal because prior single-degradation methods often fail on both degradations simultaneously.
- Raindrops and camera-side reflections frequently co-occur in rainy autonomous-driving scenes, challenging onboard visual recognition and vehicle cameras.
- Reduced visibility from adherent raindrops and reflections may lead to severe driving safety hazards.
- UR3 seeks a unified model for raindrop-reflection composite degradation, whereas earlier methods treated raindrop and reflection removal separately.
- The First Challenge was organized at the Second LoViF Workshop at ECCV 2026 to find a practical simultaneous-removal solution.
- The challenge is intended to support applications including autonomous driving, photography, and video surveillance.
2 UR3 Challenge
The UR3 Challenge established a first competition for unified raindrop and reflection removal using the real-shot RDRF dataset. It defined a weighted PSNR, SSIM, and LPIPS evaluation protocol across development and testing phases.
- The UR3 Challenge was the first competition organized to advance unified removal of raindrops and reflections.
- 2.1 RDRF dataset: The RDRF dataset contains 307 unique scenes, divided into 216 training scenes with 9003 image pairs and 91 validation scenes with 277 image pairs.
- 2.1 RDRF dataset: An additional testing set contains 100 image pairs from 55 scenes, with only degraded images publicly released for fairness.
- 2.2 Evaluation protocol: Ranking uses PSNR, SSIM, and LPIPS, with Score = PSNR(Y) + 10×SSIM(Y) − 5×LPIPS.
- 2.2 Evaluation protocol: PSNR and SSIM are evaluated on the Y channel, while LPIPS measures normalized perceptual distance using AlexNet.
- 2.3 Challenge phases: The development phase received 1176 submissions from 56 participants, while the testing phase received 406 submissions from 38 participants.
3 Challenge Results
Xedit Master achieved the best overall challenge performance, while the authors identified multi-frame fusion as inconsistent with the intended single-image setting. Strict single-frame retesting changed the ranking, placing ACVLAB third after removing multi-frame fusion.
- 41.8226: Xedit Master achieved the best overall final score and the highest PSNR of 32.3993.
- 41.7405: tysl ranked second overall and placed second across PSNR, SSIM, and LPIPS.
- 41.1416: GIST-IVL ranked third overall and achieved the highest SSIM of 0.9680 and LPIPS of 0.0388.
- Multi-frame fusion can enhance restoration because one ground-truth image may correspond to multiple degraded images, but it conflicts with the challenge’s intended single-image task.
- Under strict single-frame retesting, ACVLAB ranked third after multi-frame fusion was removed.
4 Teams and Methods
The teams use distinct two-stage restoration designs, including cascaded refinement and diffusion-guided priors, to jointly remove raindrops and reflections. Their pipelines combine degradation removal with refinement or prior-guided fidelity correction.
- Xedit Master: Xedit Master uses a two-stage framework with an RDNet remover followed by a NAFNet refiner for residual-artifact correction and detail recovery.The remover jointly suppresses raindrops and reflections, while the refiner targets remaining artifacts and fine details.
- tysl: tysl adapts GenSIRR with LoRA to generate a clean-image restoration prior, then refines it using the original degraded image.The diffusion output is treated as a prior because it may contain local geometric or color inconsistencies.
- tysl: The tysl refiner aligns diffusion-prior and degraded-image features through bidirectional deformable attention to correct local inconsistencies.A ConvNeXt-based estimator predicts a spatial prompt map for adaptive refinement-feature modulation.
4.3 GIST-IVL
GIST-IVL proposes MAFNet-UR3, a three-stage framework that progressively removes raindrops, removes reflections using complementary priors, and performs final refinement. Training and inference follow this staged decomposition.
- Framework: MAFNet-UR3 uses three stages to perform unified raindrop and reflection removal from multiple observations of the same scene.The framework is illustrated as the overall GIST-IVL method.
- Stage I: Stage I fine-tunes WeatherRemover on raindrop-free targets that retain reflections, producing an intermediate result for reflection removal.At inference, Stage I suppresses raindrops while preserving reflection cues for subsequent processing.
- Stage II: Stage II adapts FUMO to the unified-removal task and conditions processing on intensity and wavelet-derived high-frequency priors.Qwen3-VL generates the intensity prior, while wavelet decomposition extracts the high-frequency prior.
- Training details: The stages are trained for 50k, 200k and 25k iterations for Stage I, the Stage II diffusion model, and its refiner, respectively.Stage III is trained for 100k iterations on 256 × 256 patches.
4.4 ACVLAB
ACVLAB introduces RDNet-SGCR, a single feed-forward network that uses semantic and geometry guidance with cascaded refinement. Its design reflects the differing spatial-frequency characteristics of raindrops and reflections.
- Framework: RDNet-SGCR is a single feed-forward restoration network that adds Semantic-and-Geometry-Guided Cascaded Refinement to RDNet.The method is presented as the ACVLAB framework in Fig. 5.
- Motivation: The method treats raindrops as sparse localized high-frequency occlusions and reflections as low-frequency semi-transparent layer-mixing artifacts.These complementary degradation characteristics motivate the cascaded refinement design.
- Feature guidance: Frozen FocalNet-L and DINOv2 encoders extract appearance and semantic features that are fused and modulated by a content-adaptive process.The supplied passage describes feature fusion and modulation but ends before specifying the full mechanism.
- Cascaded refinement: Four reversible subnetworks progressively estimate restoration residuals, with the last three outputs supervised as refinement stages.The geometry adapter is zero-initialized to preserve the RDNet initialization.
4.5 SNU-ISPL
SNU-ISPL develops DINOClear as a data-centric adaptation of DINOLight and also evaluates an ensemble of Restormer-family predictions. The approach emphasizes self-supervised semantic and structural priors, while the ensemble fuses frequency-separated outputs.
- DINOClear: DINOClear adapts DINOLight hierarchically for unified raindrop and reflection removal without increasing architectural complexity or model capacity.Its generalization strategy relies on a sequential three-phase training curriculum.
- DINOClear: DINOv2 self-supervised features provide semantic and fine-grained structural priors for distinguishing scene content from degradation artifacts.An Adaptive Feature Fusion Module combines shallow geometric features with deeper semantic representations.
- Ensemble framework: The ensemble framework combines eight Restormer-family predictions through frequency-split fusion after scene-disjoint validation selects Restormer and X-Restormer as primary backbones.The compared architectures include Restormer, X-Restormer, DRSformer, InstructIR, MambaIR and DAI.
- Efficiency and fusion: X-Restormer and Restormer require 1.16 and 0.67 seconds per image without TTA, increasing to 3.60 and 2.35 seconds with four-way TTA.The final ensemble averages low-frequency components and fuses high-frequency residuals.
4.7 AIIA-Lab
AIIA-Lab uses multi-stage restoration pipelines that combine transformer backbones, checkpoint aggregation, and refinement or self-ensemble strategies for joint raindrop and reflection removal.
- Method: The proposed pipeline adopts MSDT as its backbone and adds scene-specific checkpoint selection, interframes fusion, and Flux refinement.The overall method is organized as a multi-stage pipeline.
- Method: High-performing MSDT checkpoints are aggregated by mean or median fusion to create pseudo ground truth for Flux fine-tuning.
- Testing: Testing uses sliding-window inference with overlapping-prediction averaging and edge padding to support arbitrary image resolutions.Recomposition handles dimensions that are not divisible by the window size.
- Method: A separate approach uses XResformer with a four-stage coarse-to-fine training schedule and self-ensemble testing.Training includes patch-size changes and combines ℓ1 with a frequency-domain loss in an intermediate stage.
4.9 sunsean
sunsean combines direct and diffusion-based restoration through a two-branch ensemble, while another framework conditions diffusion restoration on degradation clusters and pseudo-references.
- Two-branch ensemble: The two-branch ensemble combines direct restoration RD with diffusion-based DAI, then linearly blends their restored predictions.The branches are described as providing complementary restoration characteristics.
- Inference: For high-resolution inputs, DAI uses 512 × 512 overlapping tiles with 64-pixel overlap before merging predictions into a full-resolution output.
- Checkpoint selection: DAI selects checkpoints by scene complexity, using one checkpoint for simple scenes and averaging two complementary checkpoints for challenging scenes.The two predictions are averaged with equal weights in challenging scenes.
- Conditional diffusion: The conditional diffusion framework is based on Stable Diffusion 2.1 and clusters DINOv3 features with Affinity Propagation for cluster-level degradation conditioning.
- Conditional diffusion: A scene-level pseudo-reference generated by pixel-wise median aggregation is concatenated with each degraded input image.
4.11 ustc_pi_lab
The described methods use multi-stage residual correction, mixture-of-experts routing, and self-ensemble inference to address raindrop and reflection degradation.
- ustc_pi_lab: The two-stage framework first uses Histoformer to produce an initial restoration, then applies a signed multi-band residual subtractor for correction.The subtractor processes a 21-channel representation containing the degraded input, initial output, their difference, and frequency features.
- ustc_pi_lab: Figure 9 presents the pipeline of the method proposed by Team ustc_pi_lab.
- ustc_pi_lab: The training objective combines image reconstruction, residual supervision, multi-band consistency, and clean-region regularization.The pretrained Histoformer checkpoint is counted as external model usage.
- Mixture of experts: The mixture-of-experts network uses a shared convolutional backbone and a spatial router that activates the top-2 of eight convolutional experts.Dual reconstruction heads predict the transmission layer T and reflection layer R under a layer-consistency loss.
- Testing: Testing restores each 1080 × 720 image with full-image inference and averages predictions from the original and horizontally flipped views.The method uses a single checkpoint without tiling.
Teams and Affiliations
The passage directs readers to supplementary material for additional information about teams and affiliations.
- Teams and Affiliations: Additional information is available in the supplementary material.
- Teams and Affiliations: The supplied passage does not list team names or affiliations.
- Teams and Affiliations: The passage provides a cross-reference rather than substantive affiliation details.