Research map / 3D vision / Face and shape
VAREN: Very Accurate and Realistic Equine Network
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.
Builds VAREN, a learned 3D articulated horse body model from ~4000 dynamic scans of 50 horses, with an anatomical skeleton and muscle-grouped pose-dependent deformations.
Problem
Existing 3D articulated animal shape models for horses are learned from toys or figurines, lack realism, and do not model pose-dependent muscle deformation or anatomically meaningful joints.
Approach
Learn a parametric horse model from real 3D scans with an anatomical skeleton and per-muscle pose-dependent deformations, where each muscle's deformation is a learned function of pose weighted by body-part influence matrix A, rather than distance-based locality as in STAR or global linear blend shapes as in SMPL.
Key results
- On the Out-shape test set, VAREN achieves Chamfer distance 10.88 mm and mesh-to-scan 7.78 mm, versus 11.03/8.36 for STAR-style deformations and 12.00/10.07 for the baseline without pose-dependent deformations.
- On the In-shape test set, VAREN reaches 9.37 mm Chamfer and 6.35 mm mesh-to-scan, the best among baseline, SMPL, and STAR.
- hSMAL yields 21.92 mm Chamfer on the Out-shape set, far worse than the learned VAREN model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pose-dependent deformations (ablation) | VAREN · In-shape testset | Chamfer distance (mm): 10.32 -> 9.37 | Helped |
| add pose-dependent deformations (ablation) | VAREN · In-shape testset | Mesh-to-scan distance (mm): 8.20 -> 6.35 | Helped |
| use muscle-based pose-dependent deformations instead of STAR deformations | VAREN · In-shape testset | Chamfer distance (mm): 9.51 -> 9.37 | Helped |
| use muscle-based pose-dependent deformations instead of STAR deformations | VAREN · In-shape testset | Mesh-to-scan distance (mm): 6.99 -> 6.35 | Helped |
| add pose-dependent deformations (ablation) | VAREN · Out-shape testset | Chamfer distance (mm): 12.00 -> 10.88 | Helped |
| add pose-dependent deformations (ablation) | VAREN · Out-shape testset | Mesh-to-scan distance (mm): 10.07 -> 7.78 | Helped |
| use muscle-based pose-dependent deformations instead of STAR deformations | VAREN · Out-shape testset | Chamfer distance (mm): 11.03 -> 10.88 | Helped |
| use muscle-based pose-dependent deformations instead of STAR deformations | VAREN · Out-shape testset | Mesh-to-scan distance (mm): 8.36 -> 7.78 | Helped |
Surprising findings
- The hSMAL toy-based model is substantially less accurate on real horse scans (21.92 mm vs 10.88 mm Chamfer) than even the model without pose-dependent deformations.
Limitations
- Ears are excluded from the pose-dependent deformation model, and head, tail, ankles and hooves are removed from error evaluation.
- Alignment relies on manual landmark annotation for prototypes and about 1% of dynamic frames, and frames with missing parts must be removed.
- Muscle activation is not measured directly; deformations are only observable muscle shape changes.
Experimental setup
Datasets: dynamic 3D horse scans captured with 3dMD setup (about 3.6K training scans)
Benchmarks: In-shape test set (275 frames, 5 horses) Out-shape test set (154 frames, 6 horses)
Metrics: Chamfer distance (mm) mesh-to-scan distance (mm)
Baselines: hSMAL SMPL pose-dependent deformations STAR pose-dependent deformations model without pose-dependent deformations
Builds on: SMPL STAR hSMAL SMAL
Closest papers by meaning
- 4DEquine: Disentangling Motion and Appearance for 4D Equine Reconstruction from Monocular Video (CVPR 2026 · similarity 0.88)
- LASSIE: Learning Articulated Shapes from Sparse Image Ensemble via 3D Part Discovery (NeurIPS 2022 · similarity 0.88)
- 3D Menagerie: Modeling the 3D Shape and Pose of Animals (CVPR 2017 · similarity 0.87)
- Learning the 3D Fauna of the Web (CVPR 2024 · similarity 0.85)
- SAOR: Single-View Articulated Object Reconstruction (CVPR 2024 · similarity 0.84)
- SAM 3D Animal: Promptable Animal 3D Reconstruction from Images in the Wild (NeurIPS 2026 · similarity 0.83)
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.