Research map / Point clouds / Point cloud classification
NeuroGauss4D-PCI: 4D Neural Fields and Gaussian Deformation Fields for Point Cloud Interpolation
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 NeuroGauss4D-PCI, which combines a 4D neural field with iterative Gaussian soft clustering, RBF-based Gaussian residuals, and a 4D Gaussian deformation field to interpolate point cloud frames on DHB and NL-Drive.
Problem
Point cloud frame interpolation from sparse temporal samples is hard because point clouds are sparse and unordered, and existing methods struggle to model non-rigid, non-linear 4D deformations over longer time spans.
Approach
Represents the point cloud as soft-clustered Gaussians whose parameters are interpolated over time with RBF kernels to form residuals, tracked by a 4D Gaussian deformation field, and fused attentively with latent features from a 4D neural field to predict the target-time point cloud.
Key results
- On DHB, overall CD 0.42e-3 and EMD 2.69 versus NeuralPCI's 0.54 and 3.68, with 0.10M parameters versus 1.85M.
- On NL-Drive, average CD 0.78 and EMD 95.95 without outlier removal (0.72 and 78.66 with it), beating NeuralPCI's 0.80 and 97.03.
- On KITTIs scene flow, EPE_3D 0.0190 versus GMSF's 0.0215, with ACC_S 97.72 and Outliers 9.43.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeuroGauss4D-PCI instead of NeuralPCI | point cloud interpolation · DHB (Overall) | CD: 0.54 -> 0.42 | Helped |
| Use NeuroGauss4D-PCI instead of NeuralPCI | point cloud interpolation · NL-Drive (Average) | EMD: 97.03 -> 78.66 | Helped |
| Add outlier removal during preprocessing | NeuroGauss4D-PCI · NL-Drive (Average) | CD: 0.78 -> 0.72 | Helped |
| Use NeuroGauss4D-PCI for scene flow estimation | KITTI scene flow · KITTIs | EPE3D: 0.0215 -> 0.0190 | Helped |
| Add Gaussian point cloud representation (ablation) | neural field-only model · DHB | CD: 0.58 -> 0.57 | Helped |
| Remove the neural field, using only Gaussian point cloud and T-RBF-GR (ablation) | NeuroGauss4D-PCI · DHB | CD: 0.42 -> 0.59 | Hurt |
| Add 4D Gaussian deformation to Gaussian point cloud representation (ablation) | Gaussian point cloud model · DHB | CD: 0.59 -> 0.50 | Helped |
| Add T-RBF-GR to Gaussian point cloud and 4D deformation (ablation) | Gaussian point cloud + 4D Deformation · DHB | CD: 0.50 -> 0.49 | Helped |
| Use direct latent-geometric concatenation instead of Fast-LG-Fusion (ablation) | NeuroGauss4D-PCI · DHB | CD: 0.44 -> 0.42 | Helped |
| Use Fast-LG-Fusion instead of direct concatenation (ablation) | NeuroGauss4D-PCI · NL-Drive | CD: 0.79 -> 0.78 | Helped |
Surprising findings
- Adding the Gaussian point cloud representation alone only slightly lowers CD (by 0.01 on DHB and 0.03 on NL-Drive), and removing the neural field severely degrades performance.
- The reported parameter count inconsistently appears as 0.09M in the text and 0.10M in the table.
Limitations
- Interpretability is limited by the opacity of deep networks, as acknowledged by the authors.
- Runtime is dominated by per-scene optimization, with about 90% of time spent on optimization and 10% on inference.
- Gaussian representation alone gives limited gains over the neural field baseline in the ablation.
Experimental setup
Datasets: DHB (Dynamic Human Bodies) NL-Drive KITTI Scene Flow (KITTIs, KITTIo)
Benchmarks: DHB NL-Drive KITTIs KITTIo
Metrics: Chamfer Distance Earth Mover's Distance EPE_3D ACC_S ACC_R Outliers
Baselines: IDEA-Net PointINet NSFP PV-RAFT NeuralPCI FlowNet3D PointPWC Bi-PointFlow PT-FlowNet GMSF SCOOP+ SPFlowNet
Builds on: 3D Gaussian Splatting DGCNN NeuralPCI
Closest papers by meaning
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.87)
- Motion2VecSets: 4D Latent Vector Set Diffusion for Non-rigid Shape Reconstruction and Tracking (CVPR 2024 · similarity 0.86)
- 4Deform: Neural Surface Deformation for Robust Shape Interpolation (CVPR 2025 · similarity 0.86)
- LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis (CVPR 2024 · similarity 0.86)
- LeanGaussian: Breaking Pixel or Point Cloud Correspondence in Modeling 3D Gaussians (CVPR 2025 · similarity 0.86)
- Dynamic 3D Gaussian Fields for Urban Areas (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.