RaindropGS: A Benchmark for 3D Gaussian Splatting under Raindrop Conditions

TL;DR

RaindropGS introduces the first end-to-end benchmark for evaluating 3D Gaussian Splatting ($3DGS$) under real raindrop contamination. It combines a real-world, multi-view dataset with three aligned image sets per scene (raindrop-focused, background-focused, and rain-free ground truth) and a three-stage pipeline: data preparation, data processing, and raindrop-aware 3DGS evaluation. The study analyzes how raindrops impact camera pose estimation, point-cloud initialization, and single-image deraining, and assesses multiple 3DGS variants with different pre-processing. Key findings show that raindrop interference degrades pose/init, and that raindrop removal helps but cannot fully bridge the gap, with GS-W offering the strongest robustness to view inconsistency. The benchmark establishes realistic evaluation for advancing robust 3D reconstruction in challenging, real-world rain conditions.

Abstract

3D Gaussian Splatting (3DGS) under raindrop conditions suffers from severe occlusions and optical distortions caused by raindrop contamination on the camera lens, substantially degrading reconstruction quality. Existing benchmarks typically evaluate 3DGS using synthetic raindrop images with known camera poses (constrained images), assuming ideal conditions. However, in real-world scenarios, raindrops often interfere with accurate camera pose estimation and point cloud initialization. Moreover, a significant domain gap between synthetic and real raindrops further impairs generalization. To tackle these issues, we introduce RaindropGS, a comprehensive benchmark designed to evaluate the full 3DGS pipeline-from unconstrained, raindrop-corrupted images to clear 3DGS reconstructions. Specifically, the whole benchmark pipeline consists of three parts: data preparation, data processing, and raindrop-aware 3DGS evaluation, including types of raindrop interference, camera pose estimation and point cloud initialization, single image rain removal comparison, and 3D Gaussian training comparison. First, we collect a real-world raindrop reconstruction dataset, in which each scene contains three aligned image sets: raindrop-focused, background-focused, and rain-free ground truth, enabling a comprehensive evaluation of reconstruction quality under different focus conditions. Through comprehensive experiments and analyses, we reveal critical insights into the performance limitations of existing 3DGS methods on unconstrained raindrop images and the varying impact of different pipeline components: the impact of camera focus position on 3DGS reconstruction performance, and the interference caused by inaccurate pose and point cloud initialization on reconstruction. These insights establish clear directions for developing more robust 3DGS methods under raindrop conditions.

Figures (4)

• Figure 1: 3DGS Raindrop Reconstruction Benchmark Pipeline. We develop the first benchmark for comprehensively evaluating 3DGS performance under raindrop conditions. The benchmark begins with real-world dataset collection, proceeds through data processing, and ends with a raindrop- aware 3DGS evaluation. In particular, we assess how raindrop-induced image contamination reduces the number of points available for cloud initialization and degrades camera pose estimation, and how these factors impact the performance of 3DGS methods.
• Figure 2: Example of existing raindrop 3D datasets (DerainNeRF li2024derainnerf, WeatherGS qian2024weathergs, DerainGS liu2025deraings) and our RaindropGS Dataset. As indicated by the red boxes, existing datasets exhibit the same raindrop distribution across different viewpoints; in contrast, the green boxes illustrate the diversity of raindrop distributions in our dataset. For each viewpoint, we include both raindrop-focused and background-focused images and provide corresponding clear images for 3DGS performance evaluation.
• Figure 3: Qualitative Comparison among 3DGS methods: On the raindrop-focused dataset, the original 3DGS loses structural integrity, while GS-W and 3DGS-MCMC retains scene completeness. WeatherGS faithfully reconstructs the scene after raindrop removal but fails to correct background blur. On the background-focused dataset, all methods show moderate performance with artifacts.
• Figure 4: Qualitative Comparison across data pre-processing: Uformer and Restormer exhibit comparable raindrop removal capability. While COLMAP, when successful in estimating camera poses, reconstructs more fine-grained details than VGGT.