Dynamic Topology Reconstruction for Deformable 3D Scenes

23 concepts21 papers1 proposal

When tracking, reconstructing, or simulating a moving 3D object—whether it is a person dancing, a garment flowing, or an organ deforming during surgery—computational models typically represent the shape as a mesh or a graph. Traditional pipelines rely on a foundational assumption: while the geometric positions of the vertices change over time, the connections between those vertices remain strictly fixed. This assumption is known as topological consistency.

However, the real world is topologically dynamic. A person takes off a jacket, a hand closes into a fist and creates new surface contact, a coat is unbuttoned, or a material fractures and splits apart. When these events occur, assuming a fixed topology forces algorithms to stretch geometry artificially across newly formed gaps or incorrectly fuse surfaces together.

To overcome this, a major research direction has emerged focused on dynamic topology reconstruction: developing representations and algorithms that can adapt to changing surface connectivity on the fly.

The Failure Mode of Fixed Connectivity

To understand why traditional methods fail, consider non-rigid registration or template-based tracking. If a template mesh assumes a human mouth is closed, the upper and lower lips are permanently connected by edges in the mesh graph. When the person opens their mouth, a fixed-topology tracker cannot break those edges; instead, it stretches them across the open gap, creating web-like geometric distortions.

As noted by Cai et al. and Li et al., these transitions violate the assumption of surface consistency, causing traditional volumetric fusion and non-rigid registration methods to suffer from severe tracking errors.

This limitation spans multiple domains:

  • Human Capture & Rigging: Template-based 3D human capture systems fail during significant shape alterations, such as when a subject removes a piece of clothing or interacts with external objects (Suo et al.; Yu et al.). Similarly, motion-aware rigging frameworks struggle if characters undergo structural changes during a sequence (Zhou et al.).
  • Surface Alignment: Non-rigid matching and registration frameworks often fail to account for topological variations when surfaces make contact or biological structures divide (Sánta et al.; Li et al.).
  • Scanning & Simulation: Single-camera scanning setups cannot accurately model scenarios where a subject removes layers (Dou; Elanattil et al.), while automated mesh-closing algorithms fail when encountering complex surgical incisions or multiple disconnected surfaces (Wang).

Mechanism 1: Implicit Neural Fields and Volumetric Fusion

The most prominent strategy for bypassing the limitations of explicit mesh connectivity is the shift toward implicit neural representations and volumetric fusion. Instead of tracking explicit vertices and edges, these methods represent shapes as continuous level sets (such as Signed Distance Fields or Neural Radiance Fields).

Because the surface is extracted implicitly at each time step, the topology can naturally split, merge, or tear without requiring explicit tracking of edge connectivity.

Fixed Mesh Topology:   [Vertex A] --- (Stretched Edge) --- [Vertex B]  --> Artifacts
Implicit Level Set:    ( Surface 1 )  ======>  ( Surface 1 )   ( Surface 2 ) --> Clean Split

For instance, Cai et al. combine neural implicit functions with an invertible deforming network to maintain cycle consistency while allowing topological changes like a mouth opening. To handle real-time tracking, Li et al. introduce a topology-change-aware volumetric fusion pipeline that explicitly detects and represents geometric discontinuities from a single RGB-D camera. By decoupling geometry from a rigid template, frameworks like Function4D (Yu et al.) can successfully reconstruct complex human actions, such as opening a bag or taking off a jacket, which would completely break traditional systems.

Mechanism 2: Dynamic Graph Updating and Remeshing

When explicit meshes or graphs must be used—such as in physics-based simulations or Graph Neural Networks (GNNs)—the solution is to update the underlying connectivity graph dynamically on the fly.

In physical simulations, treating cracks or cuts as simple broken links in a static network fails to capture the geometric reality of a material splitting apart. Hodgins et al. pioneered a dynamic remeshing technique for brittle fracture that cuts edges and instantiates new surfaces along stress lines. In modern reduced-order modeling, Chen et al. represent deformation bases as continuous neural fields (LiCROM), allowing real-time simulations to adapt instantly to topological changes like hole punching or cutting without requiring model retraining.

For real-time motion estimation, static graphs cause connections to become permanently and incorrectly fused when two surfaces touch. To solve this, Lin et al. utilize a GNN combined with LSTM units (OcclusionFusion) to dynamically predict and update graph connections, allowing the model to track surfaces accurately even as they touch and subsequently separate.

Mechanism 3: Intrinsic Alignment and State Dictionaries

A third approach treats topological changes either as transitions between discrete geometric states or optimizes for properties that remain invariant to connectivity changes.

Early work by Tung et al. managed the massive data demands of 3D video by clustering frames into distinct poses using Reeb graphs, modeling topological transitions explicitly via a Markov network. This allowed the system to handle changing contact points (like hands touching hips) by treating them as predictable state transitions.

For non-rigid alignment of complex objects like garments, Lin et al. separate the problem into an extrinsic fitting stage and a non-linear intrinsic refinement stage. This allows the pipeline to align surfaces effectively even when the underlying manifold structure is altered by actions like buttoning or unbuttoning. Similarly, in neural pose transfer, Wang et al. bypass point-to-point correspondences entirely by adapting image style transfer techniques, allowing poses to be mapped across shapes with completely different vertex orders and topologies.

Key Foundational Work

Topology dictionary with Markov model for 3D video content-based skimming and description (2009) — contributed 1 concept elements to this direction.
Complete Dense Stereovision Using Level Set Methods (1998) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_connectivity_constraints_6acdbb41,
  title = {dynamic connectivity constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-connectivity-constraints/rdr_6acdbb41ec3c4a1fa7840a66ddb01b4e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_dynamic_connectivity_constraints_6acdbb41_v1,
  title = {dynamic connectivity constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-connectivity-constraints/rdr_6acdbb41ec3c4a1fa7840a66ddb01b4e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake