3DGS-W: Efficient Dense Grid Encoding for 3D Gaussian Splatting in the Wild

Abstract

Novel view synthesis from unconstrained outdoor photo collections remains challenging for robust 3D scene representation. 3D Gaussian Splatting (3DGS) achieves real-time rendering with high visual fidelity, yet its static formulation is vulnerable to illumination changes, weather variations, and transient occluders. Optimizing these factors directly into Gaussian attributes degrades reconstruction quality, producing blur artifacts and distorting static geometry. NeRF-based methods address this through appearance embeddings and visibility modeling, but their implicit volume rendering incurs substantial training and rendering overhead.In this paper, we present 3DGS-W, a robust outdoor 3DGS framework that explicitly decouples view-specific appearance compensation from transient visibility estimation. An image-conditioned residual appearance module compensates for photometric variations without altering the underlying static geometry. A complementary 2D multi-resolution Dense Grid visibility module predicts pixel-wise masks in the image domain, directly modeling the view-dependent nature of transient occlusions. By offloading transient modeling from 3D Gaussian attributes to an efficient 2D dense-grid lookup, our method suppresses dynamic interference with low per-query compute overhead; the scene-linear storage cost of its per-image grids is quantified in the experiments. Across three Phototourism landmarks, 3DGS-W obtains the best average PSNR and LPIPS among the compared methods; however, SWAG retains higher PSNR and SSIM on Trevi Fountain. In the specified visibility-module benchmark, the Dense Grid design processes 320.85 iterations per second, approximately 5.2 times the throughput of the Ha-NeRF-style positional-encoding baseline.