Research map / 3D vision / Gaussian splatting

AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting

NeurIPS 2026 · 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 AdpSplit, an error-driven adaptive split operator for 3DGS that sets the number of children from L1-error regions and initializes them from region statistics, cutting training time of accelerated 3DGS pipelines by 9.2-22.3% as a drop-in replacement.

Problem

Vanilla 3DGS adaptive density control uses fixed-cardinality random splitting, which needs many densification rounds to expose fine structure; shortening densification schedules to save time degrades reconstruction quality.

Approach

Replace fixed binary random splitting with an error-driven split: the number of children equals the number of attributable high-L1-error regions dominated by the parent, each child is initialized from the region's centroid (back-projected at the parent's Mahalanobis-closest depth), covariance, and ground-truth color, and cross-view duplicates are merged.

Key results

• With FastGS, AdpSplit matches full-schedule PSNR on MipNeRF360 (27.55 vs 27.54) while cutting training time by 16.4% (171s to 143s).
• 3DGS-AdpSplit restores MipNeRF360 PSNR to 27.54 (vs 27.41 for shortened 3DGS-S) with 23.0% training time reduction versus 3DGS.
• AdpSplit reduces training time by 9.2%-22.3% across accelerated pipelines as a drop-in split replacement, per the abstract.

Surprising findings

• Naively increasing the number of random split children (5 children with shortened schedule) yields more Gaussians but lower PSNR than the binary-split baseline (27.33 vs 27.53).

Limitations

• Effective mainly under limited (9k-iteration) densification; more elaborate splitting may reduce densification steps further.
• Early-stage error-driven initialization can make overly large Gaussians produce floating children from unreliable error regions.
• Some cases (DashGaussian-AdpSplit) take longer than the shortened baseline due to deep cross-view merging graphs; no error bars reported.

Experimental setup

Datasets: MipNeRF360 Deep-Blending Tanks&Temples

Benchmarks: MipNeRF360 Deep-Blending Tanks&Temples

Metrics: PSNR SSIM LPIPS training time FPS number of Gaussians peak GPU memory

Baselines: 3DGS MiniSplatting DashGaussian SpeedySplat FastGS

Builds on: 3D Gaussian Splatting adaptive density control compact-box rasterization

Closest papers by meaning

• GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
• Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
• Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
• ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.91)
• FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.90)
• DC4GS: Directional Consistency-Driven Adaptive Density Control for 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)