Source-linked AI summary
Dropping Anchor and Spherical Harmonics for Sparse-view Gaussian Splatting
Shuangkang Fang, I-Chao Shen, Xuanyang Zhang, Zesheng Wang, Yufeng Wang, Wenrui Ding, Gang Yu, Takeo Igarashi
TL;DR
Sparse-view 3DGS overfits because independent Gaussian Dropout can be neutralized by neighboring Gaussians, and high-degree spherical harmonics also contribute to overfitting. DropAnSH-GS drops anchor-centered Gaussian neighborhoods and high-degree SH, outperforming existing Dropout techniques with negligible computational cost while supporting post-training compression. Its anchor sampling and neighbor selection remain simple scope boundaries for future improvement.
Problem
Independent Gaussian Dropout is weakened by local spatial compensation, while high-degree spherical harmonics contribute to sparse-view overfitting.
Method
DropAnSH-GS drops spatially related Gaussian clusters around randomly selected anchors and applies Dropout to high-degree spherical harmonic coefficients.
Results
DropAnSH-GS consistently outperforms existing Dropout techniques across sparse-view datasets and 3DGS variants with negligible additional computational cost.
Takeaways & Limitations
The method strengthens sparse-view regularization, supports more robust scene representations, and enables post-training model compression through high-degree SH pruning.
Takeaways & Limitations
Uniform random anchor sampling and Euclidean-distance neighbor selection may not suit non-uniform Gaussian importance, anisotropic Gaussians, or local scene differences.
Abstract
from arXiv · showhide
Recent 3D Gaussian Splatting (3DGS) Dropout methods address overfitting under sparse-view conditions by randomly nullifying Gaussian opacities. However, we identify a neighbor compensation effect in these approaches: dropped Gaussians are often compensated by their neighbors, weakening the intended regularization. Moreover, these methods overlook the contribution of high-degree spherical harmonic coefficients (SH) to overfitting. To address these issues, we propose DropAnSH-GS, a novel anchor-based Dropout strategy. Rather than dropping Gaussians independently, our method randomly selects certain Gaussians as anchors and simultaneously removes their spatial neighbors. This effectively disrupts local redundancies near anchors and encourages the model to learn more robust, globally informed representations. Furthermore, we extend the Dropout to color attributes by randomly dropping higher-degree SH to concentrate appearance information in lower-degree SH. This strategy further mitigates overfitting and enables flexible post-training model compression via SH truncation. Experimental results demonstrate that DropAnSH-GS substantially outperforms existing Dropout methods with negligible computational overhead, and can be readily integrated into various 3DGS variants to enhance their performances. Project Website: https://sk-fun.fun/DropAnSH-GS
1. Introduction
Sparse-view 3DGS suffers from overfitting because local Gaussian redundancy weakens independent Dropout, while high-degree spherical harmonics introduce additional overfitting. DropAnSH-GS addresses both issues by dropping spatial neighborhoods around anchors and applying Dropout to high-degree SH.
- Sparse-view training induces severe 3DGS overfitting, producing artifacts, blurring, and geometric distortions.
- Neighboring Gaussians can compensate for individually dropped Gaussians because local opacity and color attributes are highly correlated.
- DropAnSH-GS selects anchor Gaussians and drops their surrounding neighbors, eliminating correlated clusters and disrupting local compensation.
- The method also drops high-degree SH, reducing appearance overfitting, prioritizing lower-degree SH, and enabling post-hoc model compression.
- High-degree spherical harmonics improve full-view performance but degrade sparse-view performance and substantially increase model size.
- Experiments report that DropAnSH-GS outperforms existing Dropout techniques with negligible additional computational cost and applies across multiple 3DGS variants.
2. Related Work
3DGS enables efficient novel view synthesis but remains vulnerable to sparse-view overfitting. Prior work applies regularization and Gaussian Dropout, whereas DropAnSH-GS strengthens regularization by dropping neighboring Gaussian clusters and spherical harmonic coefficients.
- 3DGS has attracted attention as an efficient explicit representation for novel view synthesis following earlier NeRF-based approaches.
- Sparse-view novel view synthesis remains difficult because NeRF and 3DGS performance depends heavily on dense input views.
- DropoutGS and DropGaussian mitigate sparse-view overfitting by randomly discarding Gaussians during training.
- DropAnSH-GS strengthens regularization by discarding neighboring Gaussian clusters and applying Dropout to spherical harmonic coefficients, while offering a performance–model-size trade-off.
3. Method
DropAnSH-GS strengthens sparse-view 3DGS regularization by dropping spatially related Gaussian groups and high-degree spherical harmonic coefficients, while preserving integration with the standard training pipeline.
- Motivation: Single-Gaussian Dropout can be weakened by neighboring Gaussians compensating for the removed contribution, producing negligible color changes and weak gradients.Overlapping Gaussians exhibit local redundancy and spatial complementarity.
- Motivation: High-degree spherical harmonic coefficients can overfit sparse-view appearance, whereas reducing SH degree can improve performance and reduce model size.Existing Dropout methods primarily manipulate opacity and overlook this attribute-level regularization opportunity.
- Anchor-based Dropout: DropAnSH-GS randomly selects anchor Gaussians and drops each anchor together with surrounding neighbors to create larger information voids and disrupt local compensation.The method encourages remaining Gaussians to use broader contextual cues and non-local scene features.
- Anchor-based Dropout: During training, each Gaussian’s opacity is multiplied by a mask before alpha blending, implementing the spatial removals without changing the standard objective.The effective opacity is used in the rendering process to enforce stronger regularization.
- Spherical Harmonics Dropout: The SH Dropout mechanism randomly retains coefficients only through a selected maximum degree and zeros higher-degree terms.The retained degree gradually increases during training, after which higher-degree coefficients can be discarded for smaller, faster models without retraining.
- Evaluation: The method is evaluated through qualitative and quantitative comparisons on LLFF, including the 3-, 6-, and 9-view settings.The supplied passages identify the evaluation datasets or quantitative values only partially; the LLFF comparison is explicitly organized by view count.
4. Experiments
Experiments evaluate DropAnSH-GS across sparse-view datasets, baselines, model compression, compatibility, computational cost, and ablations. The method consistently improves rendering quality, supports SH-based compression and integration with other 3DGS variants, and adds little training overhead.
- Experimental setup: Experiments use LLFF, MipNeRF-360, and Blender with varying sparse-view settings, reporting PSNR, SSIM, and LPIPS against NeRF- and 3DGS-based baselines.Randomized methods are averaged over three independent runs.
- Comparison results: DropAnSH-GS consistently outperforms all baselines, including individual-Gaussian Dropout methods, especially on LLFF with three input views.The advantage persists as the number of input views increases and is also observed on MipNeRF-360 and Blender.
- Comparison results: Qualitative comparisons show more complete scene context, preserved structural detail, smoother geometry, and fewer Gaussian-shaped artifacts than baselines.The anchor strategy removes entire regions, discouraging overfitting to local high-opacity Gaussians.
- Model compression: Retaining only zeroth-degree SH coefficients yields better performance than vanilla 3DGS while requiring 25% of the parameters.SH Dropout concentrates appearance information in lower-degree coefficients and enables post-training truncation.
- Compatibility: Applying the framework to FSGS, CoR-GS, DNGaussian, and Scaffold-GS significantly improves their sparse-view performance.The framework is described as lightweight and modular for integration with other 3DGS variants.
- Efficiency: Training time increases by less than 2.8% compared with 3DGS, alongside an average PSNR gain of 2 dB.The added computation is nearest-neighbor search for anchor Gaussians, implemented with CUDA on GPUs.
- Ablation studies: Ablations show performance declines when either Drop Anchor or Drop SH is removed, while degree-wise SH dropping outperforms random coefficient dropping.Sensitivity analysis examines anchor probability, neighbor count, and SH Dropout probability.
5. Conclusions
The paper addresses sparse-view 3DGS overfitting by identifying weaknesses in independent Gaussian Dropout and high-degree SH usage. DropAnSH-GS removes spatial clusters and drops high-degree SH, improving 3DGS and its variants while supporting post-training scalability.
- Conclusions: Independent Gaussian Dropout is weakened by local compensation, while high-degree spherical harmonic coefficients contribute to sparse-view overfitting.These observations motivate structured spatial Dropout and appearance-attribute regularization.
- Conclusions: DropAnSH-GS discards entire spatial regions and SH attributes, disrupting local dependencies and encouraging more robust global scene representations.The method extends beyond isolated Gaussian removal to regularize both spatial and appearance information.
- Conclusions: The method significantly enhances 3DGS and its variants under limited-view conditions while enabling post-training scalability.The conclusion characterizes the approach as simple and effective.
Gaussian Splatting
Additional experiments and analyses further validate DropAnSH-GS across SH ablations, reconstructed-scene visualizations, quantitative comparisons, and future-work discussion.
- Additional experimental results: The supplementary material adds SH-dropout ablations, reconstructed 3D Gaussian visualizations, quantitative comparisons on MipNeRF-360 and Blender, and discussion of future work.These experiments are presented as further validation of effectiveness and robustness.
6. Additional Experimental Results
Additional results examine SH truncation and explain how spatial Dropout improves scene coherence under sparse views. They report stronger rendering quality and more compact representations than relevant baselines.
- SH truncation: Drop SH concentrates effective parameters in lower-degree SH coefficients, mitigating performance degradation from direct high-degree SH truncation.This supports post-training reduction of appearance representation complexity.
- SH truncation: On Blender with 8 input views, models trained with Drop SH are evaluated after truncating SH coefficients to varying degrees.The comparison includes models trained with and without the Drop SH strategy.
- Spatial regularization: Dropping anchors with their spatial neighbors creates larger information voids, encouraging broader scene context and more coherent novel-view structures.Baseline methods instead tend to rely on local information, which can produce local artifacts under limited visual input.
- Additional comparisons: Supplementary quantitative results consistently show higher rendering quality and more compact model representations than existing sparse-view 3DGS variants.These results complement the main-text quantitative comparisons.
7. Discussion on Future Work
Future work focuses on making DropAnSH-GS’s anchor and neighbor selection more adaptive to scene structure and Gaussian importance. The paper also reports more complete and natural reconstructions, while the supplied tables define SH truncation comparisons for two datasets.
- 7. Discussion on Future Work: Uniform random anchor sampling may overlook that Gaussian distributions and importance vary across scene regions.The paper suggests gradient magnitude or opacity as potentially more effective selection signals.
- 7. Discussion on Future Work: Euclidean distance may select spatially adjacent Gaussians that are not functionally complementary because Gaussians are anisotropic and scenes vary locally.Future neighbor-selection schemes could incorporate Gaussian attributes, scene characteristics, or view-dependent information.
- 7. Discussion on Future Work: Figure 8 reports that the method produces more complete and natural results for the entire reconstruction scene.
- 7. Discussion on Future Work: Tables 9 and 10 compare quantitative results on MipNeRF-360 with 12 views and Blender with 8 views using SH-degree truncation at inference.SHn denotes retaining only the first n degrees of SH coefficients.