Research map / Point clouds / Point cloud classification

RigPAPR: Rig-Based Animation of Static Neural Point Clouds from a Single-View Video

NeurIPS 2026 · 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.

Rigs a static PAPR point cloud via a transient mesh proxy and auto-rigger, then drives it with direct LBS from a single fixed-viewpoint video using two-phase depth- and track-supervised optimization.

Problem

Animating static neural point reconstructions with skeletal rigs from a single fixed-viewpoint driving video (captured or I2V-generated) without joint-boundary artifacts that plague per-splat Gaussian LBS methods.

Approach

Use PAPR, a covariance-free interpolation-based point representation whose pixels are recomposed at render time from deformed neighbours, so direct LBS on points yields seamless joints without per-primitive shape or pose-dependent correction MLPs.

Key results

• On held-out synthetic novel views, RigPAPR reaches 23.63 dB PSNR vs 20.51 dB for Mani-GS and 19.27 dB for Puppeteer.
• On the synthetic training view RigPAPR attains 29.07 dB PSNR, best among methods, while Mani-GS is 28.53 dB.
• Phase 2 improves training-view PSNR by +0.9 dB for RigPAPR and +1.8 dB for Mani-GS on synthetic scenes.

Surprising findings

• Phase 2 per-splat fine-tuning causes Mani-GS to overfit off-trajectory, producing spiky artifacts on novel poses, while Phase 1 avoids them but still shows joint-boundary artifacts.

Limitations

• Errors in monocular depth and 2D tracking priors propagate into recovered articulation and cannot be corrected by the single-view rendering loss.
• Auto-rigging quality is load-bearing; poor skeletons seed suboptimal optima, especially for unseen morphologies.
• Regions occluded in the binding frame lack appearance supervision and can render as holes under articulation.

Experimental setup

Datasets: Objaverse (synthetic subjects) captured real subjects (robot, statue) Kling 3.0 generated driving videos

Benchmarks: Objaverse synthetic subjects (T-Rex, Simpsons, fox, wolf, spider) real robot and statue captures

Metrics: PSNR SSIM LPIPS chamfer distance

Baselines: Puppeteer Mani-GS SuGaR

Builds on: PAPR Puppeteer Mani-GS CoTracker3 Depth Anything 3 OffsetOPT

Closest papers by meaning

• 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.87)
• HoliGS: Holistic Gaussian Splatting for Embodied View Synthesis (NeurIPS 2025 · similarity 0.87)
• COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.86)
• Novel View Synthesis from A Few Glimpses via Test-Time Natural Video Completion (NeurIPS 2025 · similarity 0.86)
• PAPR in Motion: Seamless Point-level 3D Scene Interpolation (CVPR 2024 · similarity 0.86)
• MorpheuS: Neural Dynamic 360deg Surface Reconstruction from Monocular RGB-D Video (CVPR 2024 · similarity 0.86)