PaperlayerView exact revision on arXiv

Source-linked AI summary

CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes

Yang Liu, Chuanchen Luo, Zhongkai Mao, Junran Peng, Zhaoxiang Zhang

arXiv:2411.00771v2cs.CV

TL;DR

Large-scale Gaussian reconstruction must preserve accurate geometry while controlling convergence, memory, storage, and training costs. CityGaussianV2 builds on 2DGS with gradient-based densification, depth regression, elongation filtering, and an optimized parallel pipeline, while introducing a geometry benchmark. The paper reports a balance of visual quality, geometric accuracy, and efficiency, including up to 10x compression and 50% lower memory usage, but evaluates under constrained lighting and faces difficulties with occluded, unobserved, and thin structures.

  • Problem

    Unstructured Gaussian primitives make accurate surface reconstruction difficult in large and complex scenes.

    Source: §1 · p. 2

  • Method

    CityGaussianV2 combines 2DGS primitives with decomposed-gradient densification, depth regression, elongation filtering, and an optimized parallel-training pipeline.

    Source: §1 · p. 2, §3.2 · pp. 5–6, §3.3 · p. 6

  • Results

    The method reports a balance of visual quality, geometric accuracy, storage, and training costs across challenging large-scale datasets.

    Source: §6 · p. 10

  • Takeaways & Limitations

    Up to 10x compression and 50% memory savings are reported for the optimized large-scale pipeline.

    Source: §1 · p. 2, §5.3 · p. 10

    Evaluation uses compensated or ideally constant lighting, while occlusion, missing observations, and TSDF fusion limit some surfaces and thin structures.

    Source: D DISCUSSION · p. 17

Abstract

from arXiv · show

Recently, 3D Gaussian Splatting (3DGS) has revolutionized radiance field reconstruction, manifesting efficient and high-fidelity novel view synthesis. However, accurately representing surfaces, especially in large and complex scenarios, remains a significant challenge due to the unstructured nature of 3DGS. In this paper, we present CityGaussianV2, a novel approach for large-scale scene reconstruction that addresses critical challenges related to geometric accuracy and efficiency. Building on the favorable generalization capabilities of 2D Gaussian Splatting (2DGS), we address its convergence and scalability issues. Specifically, we implement a decomposed-gradient-based densification and depth regression technique to eliminate blurry artifacts and accelerate convergence. To scale up, we introduce an elongation filter that mitigates Gaussian count explosion caused by 2DGS degeneration. Furthermore, we optimize the CityGaussian pipeline for parallel training, achieving up to 10$\times$ compression, at least 25% savings in training time, and a 50% decrease in memory usage. We also established standard geometry benchmarks under large-scale scenes. Experimental results demonstrate that our method strikes a promising balance between visual quality, geometric accuracy, as well as storage and training costs. The project page is available at https://dekuliutesla.github.io/CityGaussianV2/.

1 INTRODUCTION

CityGaussianV2 targets the scalability, convergence, geometric-accuracy, and evaluation challenges that limit Gaussian-based reconstruction in large scenes. It combines 2DGS optimization, efficient parallel training and compression, and a standardized geometry benchmark.

Motivation

  • 2DGS-based methods face blurred Gaussians, Gaussian-count explosion during parallel training, and unstable geometry evaluation in underobserved boundary regions.These issues limit convergence, scalability, and objective comparison in large-scale scenes.Source: §1 · p. 2

Contributions

  • The pipeline reduces training and storage costs while enabling real-time rendering through optimized parallel training, compression, and pruning.These design goals directly address the high memory, storage, and training costs reported for existing large-scale methods.Source: §1 · p. 2
  • CityGaussianV2 uses decomposed-gradient densification, depth regression, and an elongation filter to improve convergence and prevent 2DGS degeneration during scaling.It also adopts block-based parallel training, degree-2 spherical harmonics, and contribution-based pruning.Source: §1 · p. 2
  • CityGaussianV2 introduces a TnT-style evaluation protocol with visibility-based cropping to provide stable geometric assessment for large, unbounded scenes.The protocol excludes underobserved regions that can otherwise distort geometry metrics.Source: §1 · pp. 2–3
  • The authors report state-of-the-art geometric quality and efficiency, with experimental validation across challenging large-scale reconstruction settings.The paper positions the method as among the first Gaussian radiance-field approaches for large-scale surface reconstruction.Source: §1 · p. 3

2 RELATED WORKS

Related work progresses from efficient novel-view synthesis and Gaussian surface reconstruction toward large-scale training systems. However, reliable geometry evaluation and scalable, accurate surface reconstruction remain open challenges.

2.1 NOVEL VIEW SYNTHESIS

  • NeRF and 3DGS methods advance novel-view synthesis through improved rendering efficiency, anti-aliasing, storage, and training structures.3DGS is prominent because of its training convergence and rendering speed.Source: §2.1 · p. 3

2.2 SURFACE RECONSTRUCTION WITH GAUSSIANS

  • Gaussian surface-reconstruction methods regularize geometry using meshes, signed distance functions, collapsed primitives, opacity fields, ray intersections, or contribution-based pruning.These approaches target unambiguous surfaces, smoothness, outlier removal, or direct mesh extraction.Source: §2.2 · p. 3
  • Existing Gaussian surface methods have mainly been validated on small scenes or single objects, leaving scaling-related performance degradation, densification stability, and training cost underexplored.This motivates studying large-scale geometric reconstruction explicitly.Source: §2.2 · p. 3

2.3 LARGE-SCALE SCENE RECONSTRUCTION

  • Large-scale Gaussian systems introduce parallel training, compression, level-of-detail rendering, distributed optimization, and optimized rendering to reduce computational costs.Examples include VastGaussian, CityGS, HierarchicalGS, OctreeGS, DoGaussian, GrendelGS, and FlashGS.Source: §2.3 · pp. 3–4
  • Geometry accuracy remains largely overlooked in large-scale reconstruction because reliable benchmarks are lacking, motivating a benchmark and algorithm for economical training and accurate geometry.CityGaussianV2 addresses this gap with an optimization mechanism comprising depth supervision, an elongation filter, and decomposed-gradient densification.Source: §2.3 · p. 4, Figure 2 · p. 4

3 METHOD

CityGaussianV2 builds a scalable large-scene reconstruction pipeline around 2D Gaussian surfels, combining depth guidance, gradient-aware densification, elongation filtering, and integrated compression. These modifications target convergence, Gaussian-count growth, and the time, memory, and storage costs of parallel training.

3.1 PRELIMINARY

  • CityGaussian pretrains a coarse model, tunes non-overlapping blocks in parallel, then applies pruning, distillation, and vectree quantization for compression.Source: §3.1 · p. 4
  • 2D Gaussian Splatting represents scenes with oriented Gaussian disks, or surfels, and supplements them with depth-distortion and normal-alignment regularization for geometry reconstruction.Source: §3.1 · pp. 4–5

3.2 OPTIMIZATION MECHANISM

  • Depth supervision uses Depth-Anything-V2 inverse-depth estimates aligned to dataset scale, with an exponentially decayed loss weight to reduce the effect of imperfect estimates.Source: §3.2 · p. 5
  • Extreme-elongation surfels can collapse into tiny projected points whose accumulated gradients trigger exponential Gaussian proliferation and out-of-memory errors during block-wise tuning.Source: §3.2 · p. 5
  • The Elongation Filter excludes surfels with low η_n from cloning and splitting, stabilizing Gaussian-count growth without compromising pretraining performance.Source: §3.2 · p. 5
  • DGD prioritizes SSIM-derived gradients for densification, while the revised pipeline starts with SH degree 2 and trims low-contribution points during block-wise tuning.The pipeline removes separate post-pruning and distillation stages; contribution-based trimming saves 50% storage and 40% memory while reducing time cost and slightly improving performance.Source: §3.2 · p. 6, §3.3 · p. 6

4 GEOMETRIC EVALUATION PROTOCOLS

The protocol standardizes geometric evaluation by extracting and uniformly sampling surface points, then estimating crop volumes from point visibility to reduce boundary effects.

  • The protocol extracts meshes for all methods and samples equal numbers of points evenly on each surface before comparison.This avoids inconsistent extraction procedures and reduces underestimation caused by large triangles.Source: §4 · p. 7
  • Crop-volume estimation removes rarely observed points by rasterizing a ground-truth-initialized 3DGS field across training views.Points below the visibility threshold are excluded before height-range and alpha-shape estimation.Source: §4 · p. 7
  • Visibility-aware cropping reduces the F1-score error bar from 0.1 to 0.003 on a 1.47km2 scene with 958 views and 31.4 million points.The procedure completes within 1 minute at 1080p on a 40G A100.Source: §4 · p. 7
  • The evaluation uses 0.35m voxel downsampling and scene-dependent distance thresholds ranging from 0.3m to 0.6m.Source: §4 · p. 7

5 EXPERIMENTS

Experiments evaluate CityGaussianV2 on five large scenes against Gaussian and NeRF-based baselines, showing stronger reconstruction quality and reduced training costs.

5.1 EXPERIMENTAL SETUP

  • The experiments cover five GauU-Scene and MatrixCity scenes spanning 0.3 km2 to 2.7 km2, each with over 4,000 training and 450 test images.Ground-truth point clouds are used exclusively for geometry evaluation.Source: §5.1 · p. 7

5.2 COMPARISON WITH SOTA METHODS

  • CityGaussianV2 outperforms 2DGS by 0.01 F1 score on GauU-Scene and gains around 11% F1-score over CityGS while sacrificing 0.24 PSNR.It also achieves better rendering and geometry at both 7K and 30K iterations without parallel tuning.Source: §5.2 · pp. 8–9
  • On MatrixCity-Aerial, CityGaussianV2 achieves the best surface quality, with an F1 score twice that of 2DGS and a significant margin over CityGaussian.On street view, it provides better geometry while maintaining rendering quality comparable to CityGS.Source: §5.2 · p. 9
  • CityGaussianV2-s reduces training time by 25% and memory usage by over 50%, while CityGaussianV2-t halves training time.The small version delivers superior geometric performance and on-par rendering quality with CityGS.Source: §5.2 · p. 9

5.3 ALBATION STUDIES

  • DGD improves 1.0 PSNR, 0.04 SSIM, and almost 0.02 F1 score, while elongated-Gaussian restriction prevents Gaussian count explosion during fine-tuning.Depth-Anything-V2 supervision also considerably enhances geometric quality.Source: §5.3 · pp. 9–10
  • Pipeline ablations show that trimming enables 50% storage and memory reduction, SH degree 2 cuts storage and memory by over 25% and improves speed by 4.2 FPS, and vectree quantization reduces storage by 75%.Using the 7,000-iteration result for pretraining reduces total training from 3 hours to 2 hours and model size below 300 MB.Source: §5.3 · p. 10

6 CONCLUSION

CityGaussianV2 addresses convergence and scaling limitations of 2DGS while reducing training costs and supporting geometric reconstruction on large-scale scenes.

  • CityGaussianV2 uses 2DGS primitives while improving convergence speed and scaling capability for large-scale surface reconstruction.Source: §6 · p. 10
  • The method adds parallel training and compression for 2DGS, achieving considerably lower training cost than CityGaussian.Source: §6 · p. 10
  • Experiments on multiple challenging datasets demonstrate the method’s efficiency, effectiveness, and robustness.Source: §6 · p. 10

A ADDITIONAL QUALITATIVE COMPARISON

Additional visual comparisons show that CityGaussianV2 produces more detailed, smoother, and more complete surfaces than competing methods in large-scale scenes.

  • Some road surfaces and building facades remain poorly reconstructed under occlusion or limited observation, while TSDF fusion misses some thin structures such as spires.Source: Figure 11 · p. 15
  • CityGaussianV2 produces qualitatively better reconstructions with more geometry details and fewer outliers than GOF.GOF captures intricate structures but is obscured by a difficult-to-remove near-ground shell.Source: A ADDITIONAL QUALITATIVE COMPARISON · pp. 14–15
  • In street-view comparisons, CityGaussianV2 reconstructs smoother and more complete building and road surfaces than SuGaR, GOF, 2DGS, and CityGaussian.Its rendering quality is on par with CityGaussian.Source: Figure 9 · p. 14, A ADDITIONAL QUALITATIVE COMPARISON · p. 15
  • At 7,000 iterations, DGD removes blurry surfels and produces higher-quality results than 2DGS, 3DGS, and flattened 3DGS.The comparison links early blurring in 2DGS and flattened 3DGS to dimension collapse.Source: Figure 10 · p. 14, A ADDITIONAL QUALITATIVE COMPARISON · p. 15

B ADDITIONAL QUANTITATIVE RESULTS

The additional experiments compare CityGaussianV2 with parallel-training and geometry-focused alternatives, showing stronger geometric quality alongside efficiency and rendering trade-offs. Ablations identify the gradient source used in the final densification design.

Rendering results

  • CityGaussianV2 significantly outperforms geometry-specific methods in rendering metrics while maintaining photometric quality comparable to CityGS.Qualitative reconstructions also contain more appearance detail.Source: B ADDITIONAL QUANTITATIVE RESULTS · p. 15

Parallel training comparison

Geometry and rendering results

Densification ablation

  • Densification relies exclusively on the SSIM-loss gradient because RGB-loss participation harms PSNR and SSIM, while normal- and depth-gradient effects remain within a 0.003 F1 error bar.The ablation uses 2DGS with the Elongation Filter as its baseline.Source: B ADDITIONAL QUANTITATIVE RESULTS · p. 16, Table 7 · p. 17

C MORE IMPLEMENTATION DETAILS

The implementation follows CityGaussian defaults for primitives, partitioning, and parallel tuning, with scene-specific learning rates, block layouts, and quantization settings.

Implementation settings

  • CityGaussianV2 uses lower tuning learning rates, longer street-view densification intervals, scene-specific block partitions, and vectree quantization with an 8192-entry codebook and 0.4 ratio.GauU-Scene uses a 0.05 SSIM threshold and the central one-third foreground range; position and scaling learning rates are reduced during fine-tuning.Source: C MORE IMPLEMENTATION DETAILS · p. 17

D DISCUSSION

The discussion identifies lighting conditions, incomplete observations, thin structures, and rasterization speed as limitations or remaining boundaries of CityGaussianV2.

Limitations

  • Evaluation uses GauU-Scene and MatrixCity with compensated or ideally constant lighting, leaving illumination variance as an adaptability boundary.The authors suggest decoupled appearance modeling as a possible response.Source: D DISCUSSION · p. 17
  • Occlusion and missing observations hinder some road and facade reconstructions, while TSDF fusion struggles with thin structures such as spires.The authors point to more efficient training and advanced mesh extraction as possible remedies.Source: D DISCUSSION · p. 17
  • Despite improving rendering speed, the compression strategy remains slower than CityGS at similar Gaussian counts, motivating rasterizer and Level of Detail optimizations.The discussion cites deeper rasterizer optimization and LoD integration as future directions.Source: D DISCUSSION · pp. 17–18

Related papers

Loading 2411.00771v2…