TL;DR
This paper introduces an inverse rendering method that reconstructs complex, fuzzy 3D objects—such as hair, fur, and textiles—from multi-view images using explicit 1D line primitives. By avoiding implicit volumes or translucent 3D Gaussians, the resulting models achieve high visual fidelity while remaining fully compatible with standard graphics pipelines, z-buffer depth testing, and physical simulations.
What Problem Does This Paper Address?
Faithfully capturing "fuzzy" or highly detailed, anisotropic real-world structures (like fur, fabric, and plants) is a long-standing challenge in computer graphics. Traditional surface meshes struggle because these objects lack locally smooth, well-defined boundaries.
To solve this, recent view-synthesis methods have shifted toward implicit volumetric representations like NeRFs or volumetric primitives like 3D Gaussian Splatting (3DGS). While these approaches achieve stunning visual quality, they abandon explicit low-dimensional geometric structures. This makes them inherently incompatible with standard graphics pipelines that rely on depth-tested (z-buffer) rasterization, reflection modeling, and physics-based animations.
The authors bridge this gap by asking: Can we represent fuzzy, weakly defined geometry using structured, explicit primitives without sacrificing visual fidelity?
Methods
The authors represent scenes using a collection of explicit 1D line segments and polylines. Each vertex contains a 3D position, learnable auxiliary attributes (such as Spherical Harmonics for view-dependent shading), and a per-edge opacity value.
To optimize these primitives directly from images, they developed a fully differentiable rasterization framework featuring three key components:
- Width-Free Bresenham Rasterization: Instead of optimizing continuous line widths (which is highly unstable at subpixel scales), the method rasterizes lines onto a $2\times$ subpixel grid using the classic Bresenham algorithm.
- Stochastic Differentiable Rasterizer: Building on DiffSoup (a stochastic rasterizer for triangles), the framework treats line opacities as existence probabilities during training. Primitives are stochastically selected based on depth, allowing gradients to propagate through discrete depth tests. To handle positional gradients along the inherently discrete Bresenham fragment boundaries, they adapt a microedge gradient method.
- Adaptive Re-lining: To dynamically optimize the 1D polyline topology, the framework periodically performs discrete structural updates: pruning low-opacity or short edges, merging nearby endpoints, and splitting long edges to efficiently distribute the vertex budget.
Datasets
The method was evaluated on two primary datasets:
- Shelly Dataset: A synthetic benchmark containing path-traced renderings of complex furry objects (e.g., a horse, a kitten, and a woolly ball) created by professional artists.
- Fuzzy Dataset: A new, publicly released real-world dataset captured by the authors. It features highly challenging geometries like cacti, fur, kiwi fruit, and textiles. Images were captured in a dark environment with a coaxial flash and polarization filters to suppress specular highlights.
Key Insights and Ideas
- Fuzziness from Opaque Aggregation: A key "aha" moment is that a semi-transparent, fuzzy appearance can emerge purely from the screen-space aggregation and anti-aliasing of many thin, opaque line segments, rather than requiring complex translucent primitives.
- Stabilizing Subpixel Gradients: Evaluating stochastic opacity gradients across an entire anti-aliasing filter kernel introduces extreme noise. The authors found that evaluating the opacity gradient only at the four fragments closest to the pixel center (the nearest neighbors) successfully stabilizes optimization.
- Topology Evolution: By allowing lines to dynamically appear, disappear, move, and reconnect via the adaptive re-lining strategy, the network can organize unstructured initializations into coherent 1D manifolds.
Key Results
- High Fidelity & Efficiency: On the synthetic Shelly dataset, the line-based approach significantly outperformed surface-based baselines (like VolSurfs) in capturing anisotropic boundaries. It achieved perceptual quality (SSIM and LPIPS) comparable to or better than volumetric methods like 3DGS, while using a highly parameter-efficient geometric representation.
- Blazing Fast Rendering: Because the output consists of standard line primitives, it integrates seamlessly into hardware-accelerated pipelines. The method achieves rendering speeds between 628 and 944 FPS on an RTX 4090, outperforming the 3DGS baseline at equivalent visual quality.
- Downstream Applications: Because the geometry is explicit, the reconstructed polylines immediately support advanced graphics workflows. The authors demonstrated direct mass-spring physics simulations running at ~55 ms/frame for 2 million vertices, physically-based hair scattering shaders, and real-time interactive previews on mobile devices.
Limitations and Future Work
While powerful, the method has a few notable limitations:
- Occluded Interiors: It cannot accurately recover heavily occluded internal structures, such as the deep interior of a thick hairstyle.
- Flat Surfaces: For objects that combine both fuzzy boundaries and large, visually flat regions (like the Woolly scene), triangle meshes remain more efficient at filling space. The authors suggest a hybrid triangle-line representation as a promising future direction.
- Scale and Level of Detail: The current formulation does not adapt line primitives across varying viewing distances, which could cause aliasing or performance issues in extreme close-ups.
Summary updated