Overcoming the Hollow Shell Problem in 3D Vision

92 concepts67 papers1 proposal

Imagine a beautifully rendered 3D model of an apple. Viewed from any angle, its skin looks photorealistic. However, if you attempt to slice it in half, simulate it dropping to the floor, or calculate its weight, the illusion shatters. In standard computer vision, the apple is completely empty.

Traditional 3D reconstruction and generation pipelines suffer from a fundamental bias: they prioritize what is visible. Because models are typically trained using 2D rendering supervision or line-of-sight sensors, representations like Neural Radiance Fields (NeRFs), 3D Gaussian Splatting, and Signed Distance Fields (SDFs) treat the world as a collection of infinitely thin, hollow shells (Hu et al.; Sanghi et al.).

To transition 3D vision from purely visual asset generation to physics-ready simulation, researchers are developing techniques to bridge this surface-to-volume gap.


Why Surface-Only Models Fail

When a 3D model lacks internal volume, it fails catastrophically in any application governed by the laws of physics.

Physical Simulation and Collapse

Physics engines operate on continuum mechanics, which require an understanding of internal mass, pressure, and connectivity. When a surface-only representation is subjected to external forces, it lacks the internal structural support to resist deformation. As Qiu et al. observe, simulating physical forces on hollow primitives causes the geometry to immediately collapse or buckle. Furthermore, if an object undergoes a destructive event, a surface-only model cannot realistically break apart; instead of solid material, a fracture reveals an empty, physically impossible void (Xu).

Property Estimation Bias

Downstream tasks like robotic grasping and autonomous driving rely on accurate spatial and material understanding. If a robot estimates an object's mass based purely on its outer shell, it cannot distinguish between a solid block and a hollow container, leading to failed grasps (Shuai et al.). In autonomous driving, because cameras cannot see through the outer panels of large vehicles like buses, occupancy networks frequently classify the vehicle's interior as empty space, violating physical reality (Zhu et al.).


Strategy 1: Internal Particle Densification

The most direct way to fix a hollow shell is to fill it. Researchers working with explicit primitives, such as 3D Gaussian Splatting, have introduced geometric filling mechanisms to populate the interiors of surface reconstructions.

Hollow Shell (Surface Only)       Internal Densification (Physics-Ready)
       _______                             _______
      /       \                           /• • • •\
     |         |            --->         |• • • • •|
      \_______/                           \• • • •/

To make visual models simulation-ready, Xie et al. implement a ray-casting procedure that detects the boundaries of a surface-only reconstruction and populates the interior with physical particles. This internal filling ensures the object preserves its volume and structural integrity during collisions.

For dynamic environments, Lei utilizes an octree-based voxel filling algorithm to generate internal particles, providing the necessary mass and structural density required to simulate wind-driven object deformations. When modeling high-impact destruction, Xu samples particles inside an object using a Signed Distance Function. Because the internal structure is pre-defined before the impact occurs, newly exposed surfaces at fracture sites appear as solid material rather than hollow fragments.


Strategy 2: Predictive Volumetric Fields

Instead of filling geometry after the fact, another paradigm trains neural networks to infer internal structures directly from surface observations by learning structural priors.

  • Multi-Layered Surfaces: Hu et al. introduce "X-Ray," a representation that treats 3D objects as a sequence of multi-layered surface scans. By training video diffusion models to generate these layers, the system reconstructs both the visible exterior and hidden internal architectures of complex shapes.
  • Anatomical Digital Twins: In human body modeling, where traditional parametric meshes are entirely hollow, Keller et al. train implicit neural representations to predict internal tissues—such as fat, muscle, and bone—solely from the external body shape, allowing for realistic tissue deformation during movement.
  • Mechanical Property Fields: Moving beyond geometry alone, Dagli et al. present feed-forward models that predict fine-grained, non-uniform material properties (such as Young’s modulus and density) throughout the entire internal volume of an asset, bridging the gap between computer vision and finite element analysis.

Strategy 3: Inward Geometric Projections and Skeletons

A third approach mathematically links boundary representations to internal volumetric structures using skeletal mid-structures.

Cohen demonstrates that volumetric solids can be constructed from surface boundaries by projecting control points inward toward a skeletal center. This mapping creates a layered trivariate spline volume that preserves the topological relationship between the surface and the interior.

By enforcing these internal structural constraints through representations like Centroidal Voronoi Tessellations (Allain et al.) or topology-preserving medial axis transforms (Wang et al.), tracking algorithms can model complex, contorted deformations. The internal skeleton acts as an anchor, preventing the physically impossible self-intersections and collapses that plague hollow geometric models.

Key Foundational Work

A framework for facial surgery simulation (2002) — contributed 2 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_surface_to_volume_representation_gap_5e9e5050,
  title = {surface-to-volume representation gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/surface-to-volume-representation-gap/rdr_5e9e5050f5c147068b9d480291698ec6},
  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_surface_to_volume_representation_gap_5e9e5050_v1,
  title = {surface-to-volume representation gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/surface-to-volume-representation-gap/rdr_5e9e5050f5c147068b9d480291698ec6/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