HiAnimal: Towards High-Fidelity and Animatable Mesh Reconstruction From Single-View In-the-Wild Animal Images.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Animal Modeling

Background:

• Estimating 3D animal shapes and poses is crucial for behavioral studies and creating realistic digital representations.
• Current single-view reconstruction methods often lack precise alignment, fine details, or consistent topology.

Purpose of the Study:

• To develop a novel framework, HiAnimal, for generating high-fidelity and animatable 3D animal models.
• To improve the accuracy, detail, and topological consistency of 3D animal reconstructions from single images.

Main Methods:

• HiAnimal integrates pixel-aligned implicit priors with a parametric model (SMAL).
• It predicts a per-pixel occupancy-UV field from an intermediate normal map for enhanced detail and generalization.
• A novel pixel-aligned UV correspondence field ensures precise alignment and consistent mesh topology.

Main Results:

• HiAnimal demonstrates superior performance compared to state-of-the-art methods in 3D animal reconstruction.
• The framework achieves high-fidelity and animatable 3D models with improved detail and geometry.
• Results show strong scaling ability to novel animal data.

Conclusions:

• HiAnimal effectively addresses limitations in single-view 3D animal reconstruction.
• The proposed method offers a robust solution for generating detailed and animatable 3D animal models.
• This advancement has significant implications for animal behavior research and virtual reality applications.