Abstract:Neural representations of 3D data have been widely adopted across various applications, particularly in recent work leveraging coordinate-based networks to model scalar or vector fields. However, these approaches face inherent challenges, such as handling thin structures and non-watertight geometries, which limit their flexibility and accuracy. In contrast, we propose a novel geometric data representation that models geometry as distributions-a powerful representation that makes no assumptions about surface genus, connectivity, or boundary conditions. Our approach uses diffusion models with a novel network architecture to learn surface point distributions, capturing fine-grained geometric details. We evaluate our representation qualitatively and quantitatively across various object types, demonstrating its effectiveness in achieving high geometric fidelity. Additionally, we explore applications using our representation, such as textured mesh representation, neural surface compression, dynamic object modeling, and rendering, highlighting its potential to advance 3D geometric learning.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is the limitations of existing 3D geometric representation methods when dealing with complex structures and non - watertight geometric shapes. Specifically:
1. **Limitations of traditional geometric representation methods**:
- Existing geometric representation methods, such as triangular meshes, voxels, point clouds, and implicit functions, face challenges when dealing with shapes with different topological structures, boundary conditions, and connectivities.
- For example, triangular meshes and polygonal meshes have inconsistent data structures when dealing with shapes with different numbers of vertices and different connectivities, and it is difficult to adapt to geometric learning tasks.
- Although voxels are suitable for learning - based tasks, they consume a huge amount of memory when capturing details at high resolutions.
- Although point clouds are easy to obtain from sensors, they are essentially samples of geometry, which may lead to information loss, and lack inherent point connectivity, making it difficult to define surface structures, boundaries, or geodesics.
- Although implicit functions are good at generating smooth surfaces and representing complex topologies, they perform poorly when modeling thin structures or non - closed geometries, and the integration with color or texture is not intuitive.
2. **The proposed new method**:
- The paper proposes a new geometric data representation method - **Geometry Distributions**, which overcomes the above problems by modeling 3D geometry as a distribution.
- This method uses diffusion models and novel network architectures to learn the distribution of surface points, can capture fine geometric details, and has no assumptions about the topology, connectivity, and boundary conditions of the surface.
- This method can handle 3D geometries with complex details, high genus, sharp features, and non - closed surfaces.
3. **Applications and advantages**:
- This method has been qualitatively and quantitatively evaluated on various object types, demonstrating its effectiveness in achieving high geometric fidelity.
- In addition, this method can also be applied to areas such as textured mesh representation, neural surface compression, dynamic object modeling, and rendering, and has broad application prospects.
In summary, this paper aims to solve the limitations of existing methods when dealing with complex and non - closed geometric shapes by introducing a new geometric representation method, thereby providing a more flexible and accurate 3D geometric representation.