Research map / Point clouds / Point cloud classification
PAPR in Motion: Seamless Point-level 3D Scene 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 point-level 3D scene interpolation between two observed scene states and PAPR in Motion, which fine-tunes a PAPR point-based renderer from start to end state with rigidity and local-averaging regularizers to produce smooth interpolations.
Problem
Synthesizing smooth, physically plausible point-level interpolations and novel views of a 3D scene observed only at two distinct states (multi-view images), without any intermediate supervision, where geometry may change substantially and non-rigidly.
Approach
Treat the start-state PAPR model as a geometry-agnostic scene template, fine-tune only point coordinates to fit the end state, and use the optimization trajectory as Lagrangian interpolation; add a local distance preservation loss and periodic local displacement averaging to keep motion cohesive, and interpolate appearance parameters by travelled-distance progress.
Key results
- PAPR in Motion achieves avg SI-FID 120.30 on synthetic scenes vs 194.64 for Dynamic Gaussian (lower is better).
- On real-world scenes, avg SI-CD 9.05 vs 37.13 for Dynamic Gaussian (x10^-3 scale).
- Ablation shows removing local displacement averaging causes part-geometry deformations and removing the rigidity loss causes excessive point drift.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PAPR in Motion instead of Dynamic Gaussian | point-level 3D scene interpolation · Synthetic scenes (Avg) | SI-FID: 194.64 -> 120.30 | Helped |
| Use PAPR in Motion instead of Dynamic Gaussian | point-level 3D scene interpolation · Real-world scenes (Avg) | SI-FID: 280.24 -> 199.74 | Helped |
| Use PAPR in Motion instead of Dynamic Gaussian | point-level 3D scene interpolation · Synthetic scenes (Avg) | SI-CD: 9.68 -> 3.94 | Helped |
| Use PAPR in Motion instead of Dynamic Gaussian | point-level 3D scene interpolation · Real-world scenes (Avg) | SI-CD: 37.13 -> 9.05 | Helped |
| Use PAPR in Motion instead of Dynamic Gaussian | point-level 3D scene interpolation · Synthetic scenes (Avg) | SI-EMD: 31.49 -> 12.62 | Helped |
| Use PAPR in Motion instead of Dynamic Gaussian | point-level 3D scene interpolation · Real-world scenes (Avg) | SI-EMD: 95.31 -> 53.02 | Helped |
| Add local displacement averaging step (LDAS) (ablation) | PAPR in Motion · Giraffe | interpolation geometry quality | Helped |
| Add local distance preserving loss (Lrigid) (ablation) | PAPR in Motion · Giraffe | interpolation geometry quality | Helped |
Limitations
- Cannot render regions visible only in the end state (e.g., newly entering objects) since the model is pretrained on the start state.
- No prior semantic knowledge, so generated motion may not replicate real-world motion.
- Evaluation relies on ground truth derived from static reconstructions at start/end states, and only one baseline is compared.
Experimental setup
Datasets: NeRF Synthetic Objaverse manually captured real-world scenes
Benchmarks: Scene Interpolation metric on synthetic and real-world scenes
Metrics: SI-FID SI-CD SI-EMD
Baselines: Dynamic Gaussian
Builds on: PAPR 3D Gaussian Splatting Dynamic Gaussian LPIPS
Closest papers by meaning
- RigPAPR: Rig-Based Animation of Static Neural Point Clouds from a Single-View Video (NeurIPS 2026 · similarity 0.86)
- Coherent 3D Scene Diffusion From a Single RGB Image (NeurIPS 2024 · similarity 0.86)
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction (CVPR 2024 · similarity 0.86)
- InstaInpaint: Instant 3D-Scene Inpainting with Masked Large Reconstruction Model (NeurIPS 2025 · similarity 0.86)
- NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis (NeurIPS 2026 · similarity 0.85)
- H3D-DGS: Exploring Heterogeneous 3D Motion Representation for Deformable 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.85)
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.