MeshLoom: Feed-Forward Non-Rigid Registration of Mesh Sequences

2026-06-15 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Expert, quick

Summary

MeshLoom is a novel feed-forward registration network designed to directly reconstruct vertex deformations across mesh sequences. This approach significantly advances non-rigid registration by overcoming limitations of existing models, which often involve costly per-instance optimization, narrow object categories, or pairwise-only inputs. MeshLoom is highly efficient, registering multiple meshes within seconds. Its core features include a topology-aware point representation that encodes the anchor mesh's topology into per-vertex features, and a multi-modal encoder that fuses this representation with frame-specific cues like shape latents and image features. A lightweight decoder then queries the resulting global motion embedding to retrieve per-vertex deformations, enabling state-of-the-art results and extending capabilities to motion interpolation and mesh morphing.

Key takeaway

For 3D graphics developers or computer vision engineers working with dynamic mesh sequences, MeshLoom offers a compelling alternative to traditional, often costly, per-instance optimization methods. You can achieve state-of-the-art non-rigid registration results rapidly, processing multiple meshes in seconds. Consider integrating this feed-forward network to enhance efficiency and unlock new capabilities like seamless motion interpolation and mesh morphing in your applications.

Key insights

MeshLoom uses a feed-forward network with topology-aware encoding and multi-modal fusion for efficient non-rigid mesh sequence registration and interpolation.

Principles

• Topology-aware encoding improves geometry understanding.
• Multi-modal fusion captures dense inter-frame correspondence.
• Global embedding enables intermediate timestamp generation.

Method

Encode anchor mesh topology into per-vertex features. Fuse with frame cues via a multi-modal encoder to create a compact global motion embedding. Query this embedding with anchor-mesh representation for per-vertex deformations.

In practice

• Registering non-rigid mesh sequences efficiently.
• Generating deformations at intermediate timestamps.
• Enabling motion interpolation and mesh morphing.

Topics

• Non-rigid Registration
• Mesh Sequences
• Feed-Forward Networks
• Topology-aware Encoding
• Motion Interpolation
• 3D Computer Vision

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective — Feed-forward 3D reconstruction offers efficient, generalizable scene modeling by directly mapping 2D inputs to 3D representations.
• ExMesh: EXplicit Mesh Reconstruction with Topology Adaptation — ExMesh directly optimizes explicit meshes by combining differentiable optimization with adaptive discrete topology updates and real-time UV maintenance.
• Point-Wise Geometry-Aware Transformer for Partial-to-Full Point Cloud Registration in Computer-Assisted Surgery — GAPR-Net fuses local and global geometry via a transformer and invariant features for robust partial-to-full point cloud registration.
• Surflo: Consistent 3D Surface Flow Model with Global State — Surflo uses a global latent state and flow matching to reconstruct arbitrary-resolution 3D surfaces from variable image inputs.
• RigPAPR: Rig-Based Animation of Static Neural Point Clouds from a Fixed-Viewpoint Video — Interpolation-based neural point rendering inherently avoids joint-boundary artifacts in rigged animation, unlike fixed-shape primitives.
• Velox: Learning Representations of 4D Geometry and Appearance — Velox learns descriptive and compressive 4D object representations from dynamic point clouds using dual decoders.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.