RigPAPR: Rig-Based Animation of Static Neural Point Clouds from a Fixed-Viewpoint Video

How to read this

Category

Auto-rigging and animation of neural point cloud reconstructions from single-view video

Contributions

• Diagnoses why direct linear blend skinning of 3D Gaussian splats produces joint-boundary gaps and spikes, tracing it to each splat carrying a fixed per-primitive shape calibrated in canonical pose that cannot bend
• Builds RigPAPR on Proximity Attention Point Rendering (PAPR), which recomposes each pixel at render time from deformed primitive positions rather than storing per-primitive shape, so the surface re-forms naturally under articulation
• Auto-rigs a static PAPR point cloud and drives it under direct LBS from a single fixed-viewpoint video, with no mesh proxy, pose-dependent correction, or category template, exceeding mesh-based and Gaussian-splatting baselines by 3+ dB PSNR at novel views

Context

The paper positions itself against the dominant animation pipelines for neural reconstructions, mesh-proxy deformation and direct Gaussian-splat LBS, and against prior rig-based neural point work that needs multi-camera or orbiting capture. It targets the more constrained but practical case of a single fixed-viewpoint driving video, the natural output of image-to-video generation.

Correctness

Backed by PSNR comparisons on synthetic subjects against mesh-based and Gaussian-splatting baselines, plus qualitative results on real subjects, but it explicitly matches (not beats) the strongest baseline at the supervised view and only pulls ahead at novel views. As of this note it is an arXiv preprint, not yet confirmed as peer reviewed.

Clarity

Standard preprint prose, clearly written, but assumes familiarity with Gaussian splatting, point-based neural rendering, and linear blend skinning jargon (PAPR, LBS, joint-boundary artifacts) without much hand-holding for newcomers to neural rendering.

How to read it

First pass: abstract and Figure 1's comparison of Gaussian splats versus the interpolation-based PAPR representation. Second pass: the introduction and the PAPR background to understand why per-primitive shape breaks under LBS. Third pass: the auto-rigging method section and the PSNR tables across synthetic and real subjects.