Research map / 3D vision / Gaussian splatting

DeRainGS: Gaussian Splatting for Enhanced Scene Reconstruction in Rainy Environments

AAAI 2025 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces the 3DRRE task and HydroViews dataset, and proposes DeRainGS, a 3DGS method that combines a frozen rainy-image enhancement network with a learned occlusion-mask predictor for scene reconstruction under rain.

Problem

3D scene reconstruction from images captured in rainy conditions, where rain streaks and raindrops reduce visibility and distort geometry; existing in-the-wild and deraining methods fail to handle continuous rain occlusions.

Approach

Combine a pretrained CNN-Transformer deraining network (frozen) with a 3DGS pipeline that learns per-image occlusion masks using frequency-based channel attention and a U-Net, excluding masked pixels from the photometric loss.

Key results

• On HydroViews raindrop scenes, DeRainGS reaches an average PSNR around 27 dB versus 17-23 dB for baselines (e.g., Bicycle 22.18 vs 12.25-13.27 for baselines).
• Ablation on HydroViews: full DeRainGS achieves 24.67 PSNR / 0.782 SSIM versus 17.56 / 0.659 for the vanilla 3DGS configuration.
• On the DerainNeRF waterdrop dataset, DeRainGS attains the best PSNR on all four scenes, but margins over baselines are small.

Measured effects reported in the paper

Surprising findings

• Naively using a deraining model as a preprocessor before reconstruction gives suboptimal results, and DerainNeRF performs worse than NeRF-W on HydroViews.
• The 3DGS baselines perform on par with NeRF baselines, so DeRainGS's gains come from the enhancement and masking modules rather than 3DGS rendering itself.

Limitations

• Relies on a pretrained deraining network that cannot fully remove complex rain distortions, and over-correction can introduce artifacts.
• Evaluation on the DerainNeRF dataset shows marginal differences since it mainly contains lightweight waterdrops.
• The DerainNeRF baseline weights were unavailable, so the baseline had to be retrained with AttGAN.

Experimental setup

Datasets: HydroViews 4K-Rain13k UAV-Rain1k MipNeRF-360 Tanks-and-Temples DerainNeRF waterdrop dataset

Benchmarks: HydroViews DerainNeRF waterdrop dataset

Metrics: PSNR SSIM LPIPS

Baselines: NeRF-W DerainNeRF WildGaussians GS-W Occlusion-free NeRF

Builds on: 3D Gaussian Splatting NeRF-W DerainNeRF AttGAN U-Net

Closest papers by meaning

• RainyGS: Efficient Rain Synthesis with Physically-Based Gaussian Splatting (CVPR 2025 · similarity 0.91)
• Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes (CVPR 2025 · similarity 0.88)
• EAP-GS: Efficient Augmentation of Pointcloud for 3D Gaussian Splatting in Few-shot Scene Reconstruction (CVPR 2025 · similarity 0.86)
• CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes (ICLR 2025 · similarity 0.86)
• EVolSplat: Efficient Volume-based Gaussian Splatting for Urban View Synthesis (CVPR 2025 · similarity 0.86)
• HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.86)