Research map / Point clouds / Point cloud classification

Toy-GS: Assembling Local Gaussians for Precisely Rendering Large-Scale Free Camera Trajectories

AAAI 2025 · Point clouds › Point cloud classification · approach: Scene flow and dynamic point clouds

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.

Toy-GS divides large-scale free camera trajectories into k-means camera/point-cloud regions, trains local Gaussians per region with multi-view (PatchMatch) constraints and position-aware adaptive control, and renders with a local-global fusion to improve quality and cut GPU memory.

Problem

Rendering large-scale free camera trajectories with irregular camera distributions and mixed scene types is hard for 3DGS, which needs large GPU memory and suffers holes and poor distant-object rendering when pruned region-wise (VastGaussian).

Approach

Adaptive pose-based k-means spatial partitioning with visibility-based camera selection, PatchMatch multi-view depth/normal refinement and a distance-scaled position-aware point adaptive control (PPAC), plus a local-global rendering rule that uses a region's local Gaussian when a new view is close to and sees mostly that region, otherwise falls back to the global Gaussian.

Key results

• On SCUTic, Ours (3 Areas) reaches 23.76 PSNR overall vs 22.57 for 3DGS and 21.97 for VastGS, using about 9.6 GB memory vs 16.4 GB for 3DGS.
• On the Free dataset, Ours (3 Areas) achieves 29.27 PSNR vs 27.92 for 3DGS and 27.90 for VastGS.
• On Tanks and Temples, Ours (3 Areas) gets 21.27 PSNR and 7444 MB memory, matching VastGS memory while exceeding 3DGS by about 0.4 dB.

Measured effects reported in the paper

Surprising findings

• Using only the global Gaussian (R_G) with 3 areas gives lower PSNR (21.71) than the single-area baseline (22.61), showing that pruning per region hurts without local-global fusion.
• Increasing areas from 5 to 7 with global rendering reduces PSNR, so the benefit of partitioning depends on the rendering strategy.

Limitations

• Local-global rendering relies on hand-set thresholds (max camera distance, 1/N visibility ratio) and a fixed number of areas.
• Gains over 3DGS on the Free dataset and Tanks and Temples are smaller than on SCUTic, and the method adds training complexity.
• Evaluation is limited to three datasets, with the authors' SCUTic dataset collected at one campus.

Experimental setup

Datasets: Free dataset (F2-NeRF) Tanks and Temples SCUTic

Benchmarks: SCUTic Tanks and Temples Free dataset

Metrics: PSNR SSIM LPIPS GPU memory

Baselines: 3DGS VastGaussian F2-NeRF NeRF++ mip-NeRF-360 Instant-NGP Plenoxels DVGO

Builds on: 3D Gaussian Splatting VastGaussian F2-NeRF PatchMatch PPAC

Closest papers by meaning

• Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.89)
• PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
• Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.88)
• Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis (CVPR 2024 · similarity 0.88)
• VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction (CVPR 2024 · similarity 0.88)
• MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting (CVPR 2025 · similarity 0.87)