Research map / 3D vision / Gaussian splatting

Graph-Guided Scene Reconstruction from Images with 3D Gaussian Splatting

ICLR 2025 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 0 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.

Proposes GraphGS, a pose-free large-scene 3D Gaussian Splatting framework that uses spatial-prior camera pairing (concentric nearest-neighbor pairing and a quadrant filter) to build a camera graph, then applies graph-guided multi-view consistency and betweenness-based adaptive sampling during 3DGS optimization.

Problem

Reconstructing high-quality large-scale outdoor 3D scenes from uncalibrated images with unknown or noisy camera poses and sparse viewpoints, where COLMAP's exhaustive matching is too slow for thousands of images.

Approach

Select informative image pairs via spatial priors from coarse relative poses (concentric nearest-neighbor pairing plus a binary quadrant filter over relative position and orientation) to form a camera graph, then use graph edge weights for a multi-view photometric consistency loss and betweenness-centrality node weights to sample views adaptively during 3DGS training.

Key results

• On Waymo, GraphGS reports PSNR 29.43, SSIM 0.899 and LPIPS 0.217, beating PVG (28.11 / 0.849 / 0.279) while running at 52 FPS versus PVG's 50.
• Structure estimation on 0.6k Waymo images cuts matching time from 62 min (COLMAP exhaustive) to 3 min and BA time from 140 min to 20 min; on 2k Mill-19 images COLMAP fails to finish within 24 hours while GraphGS completes in about 12 min matching and 194 min BA.
• Adaptive sampling reduces training from 30000 to 16830 iterations and from 54 min to 28 min, with PSNR rising from 28.81 to 30.36 in the ablation.

Measured effects reported in the paper

Surprising findings

• Using a deliberately worse target camera for the consistency loss is reported to improve reconstruction, which the authors liken to data augmentation.

Limitations

• Limited gains on object-centric datasets with precise poses and dense 360-degree coverage, where node importance weights become nearly uniform.
• Depends on the accuracy of coarse spatial priors; inaccurate initial xy camera distribution can cause bundle adjustment failure.
• Feature-point matching may fail in scenes with many dynamic objects; rear views and distant views remain difficult.

Experimental setup

Datasets: Waymo KITTI Mill-19

Benchmarks: Waymo KITTI Mill-19

Metrics: PSNR SSIM LPIPS FPS

Baselines: Mip-NeRF 360 S-NeRF StreetSurf Zip-NeRF UC-NeRF BARF SPARF UP-NeRF EmerNeRF 3DGS PVG Mega-NeRF

Builds on: 3D Gaussian Splatting COLMAP PlenOctrees PVG

Closest papers by meaning

• KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences (AAAI 2025 · similarity 0.90)
• Metropolis-Hastings Sampling for 3D Gaussian Reconstruction (NeurIPS 2025 · similarity 0.90)
• DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (NeurIPS 2024 · similarity 0.90)
• Holistic Large-Scale Scene Reconstruction via Mixed Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
• Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.90)
• IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.90)