Research map / 3D vision / Face and shape
TAGA: Self-supervised Learning for Template-free Animatable Gaussian Articulated Model
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 TAGA, a template-free 3D Gaussian framework that detects "ambiguous Gaussians" with incorrect skinning via a bone-based GMM and corrects them with KNN, then enforces cycle-consistency losses to refine canonical geometry and skinning from sparse pose data.
Problem
Template-free animatable articulated model reconstruction from sparse pose data suffers from canonical ambiguity, where multiple canonical models explain the same observed views, yielding poor animation quality.
Approach
Treat canonical ambiguity as Gaussians with incorrect skinning weights; detect them as point-cloud anomalies using a bone-based GMM with JSD confidence, correct their skinning via KNN neighbors, and map corrected Gaussians back to canonical space via inverse LBS as soft cycle-consistency constraints.
Key results
- On ZJU-MoCap, TAGA reaches 31.22 PSNR, 0.977 SSIM, 29.21 LPIPS* novel view, beating template-free NPC (30.76 PSNR) and HumanNeRF (30.66 PSNR).
- TAGA trains in 0.5 hours and renders at 140 FPS, versus 10 hours and 0.3 FPS for HumanNeRF on ZJU-MoCap.
- Ablation on male-3-casual (AIST++ novel poses) shows full model PSNR 28.87 vs 24.70 when removing geometry and skinning cycle losses.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Lgeo and Lskin cycle-consistency losses (ablation) | TAGA · AIST++, male3-casual, novel pose | PSNR: 24.70 -> 28.87 | Helped |
| add Lskin skinning-consistency loss (ablation) | TAGA · AIST++, male3-casual, novel pose | PSNR: 27.22 -> 28.87 | Helped |
| add soft constraints instead of directly replacing canonical Gaussians (ablation) | TAGA · AIST++, male3-casual, novel pose | PSNR: 25.11 -> 28.87 | Helped |
| use TAGA instead of HumanNeRF | PeopleSnapshot, male3-casual · PeopleSnapshot, male3-casual | PSNR: 26.13 -> 29.69 | Helped |
| use TAGA instead of HumanNeRF | PeopleSnapshot, male3-casual · PeopleSnapshot, male3-casual | LPIPS*: 27.7 -> 18.2 | Helped |
| use TAGA instead of NPC | ZJU-MoCap · ZJU-MoCap, novel view | PSNR: 30.76 -> 31.22 | Helped |
| use TAGA instead of NPC | ZJU-MoCap · ZJU-MoCap, novel view | LPIPS*: 30.84 -> 29.21 | Helped |
Limitations
- Not explicitly optimized for complex non-rigid deformations such as loose cloth; bone-based GMM assumption may not hold.
- Small holes persist in self-contact regions (e.g., under armpits) that are invisible in training data.
Experimental setup
Datasets: PeopleSnapshot ZJU-MoCap AIST++
Benchmarks: PeopleSnapshot ZJU-MoCap AIST++
Metrics: PSNR SSIM LPIPS* FPS training time
Baselines: HumanNeRF TAVA NPC InstantAvatar Anim-NeRF InstantNVR GART 3DGS-Avatar
Builds on: 3D Gaussian splatting HumanNeRF GART
Closest papers by meaning
- ArticulatedGS: Self-supervised Digital Twin Modeling of Articulated Objects using 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Building Interactable Replicas of Complex Articulated Objects via Gaussian Splatting (ICLR 2025 · similarity 0.87)
- RigGS: Rigging of 3D Gaussians for Modeling Articulated Objects in Videos (CVPR 2025 · similarity 0.86)
- Gaussian Head & Shoulders: High Fidelity Neural Upper Body Avatars with Anchor Gaussian Guided Texture Warping (ICLR 2025 · similarity 0.86)
- Emergence of a Shared Canonical Object Frame from In-the-Wild Videos (NeurIPS 2026 · similarity 0.86)
- AniGS: Animatable Gaussian Avatar from a Single Image with Inconsistent Gaussian Reconstruction (CVPR 2025 · 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.