Date: June 2024.
Source: CVPR (Computer Vision and Pattern Recognition) Conference, Seattle.
Research Summary: Data-driven 3D parametric shape models have become widely used for analyzing and generating synthetic humans, but that approach doesn’t scale easily to animals, given the sheer number of species and the difficulty of accessing live subjects to scan in 3D. VAREN argues that for a domestically important species like the horse, investing in a large real-world 3D scan dataset is worthwhile, and introduces the first 3D articulated parametric shape model learned from real horse scans. Unlike previous animal body models, VAREN operates at two resolutions and includes an anatomically grounded skeleton, with joint locations defined by real horse anatomy rather than by optimization alone. Its key innovation is a learned, interpretable pose-dependent deformation model tied to the horse’s superficial muscles rather than to distance from nearby joints, allowing the model to represent realistic muscle deformation during movement. Experiments show this muscle-based formulation outperforms deformation strategies borrowed from human body models such as SMPL and STAR, while remaining more compact and enabling analysis of the relationship between articulation and muscle deformation.
3dMD’s Role: VAREN’s entire training dataset was captured on a dynamic-3D/4D 3dMD scanning system, capturing real-world 3D scans of horses over time. The dataset spans 50 horses of varying breeds and sizes, from small ponies to large horses with more than 1.5 meters of height variation, yielding approximately 4,000 cleaned real-world 3D scans with accurately registered meshes, roughly 3,600 of which are retained for training after quality filtering. The 3dMD system captured horses both in a static neutral pose, used to build VAREN’s shape space, and performing motions such as standing, moving the neck, and moving the legs, used to train the model’s pose-dependent muscle deformation network.
Article: VAREN: Very Accurate and Realistic Equine Network.
Authors: Silvia Zuffi, Ylva Mellbin, Ci Li, Markus Hoeschle, Hedvig Kjellstrom, Senya Polikovsky, Elin Hernlund, and Michael J Black, Max Planck Institute for Intelligent Systems, Tubingen, Germany.