Introduction
In the field of 3D computer vision and graphics, representing shapes for machine learning has long been a trade-off between compactness and expressiveness. While implicit representations like Signed Distance Functions (SDFs) offer continuity and handle complex topologies, they are computationally expensive and difficult to edit locally. Conversely, explicit representations like point clouds or meshes provide direct surface data but often lack connectivity or struggle with varying topologies.
MASH (Masked Anchored SpHerical Distances) introduces a novel paradigm that bridges this gap. By treating 3D objects as a collection of local surface patches observed from a set of "anchor" viewpoints, MASH provides a differentiable, highly expressive, and compact format. This representation accounts for the "perceptual gap" in traditional methods—the fact that we typically observe shapes as a series of visible local patches rather than a single global function.
The MASH Representation
The core of the MASH framework is built upon three fundamental components that allow it to capture intricate geometry while remaining computationally efficient:
1. Anchors
The system utilizes a set of fixed points (typically around 400) distributed around the shape. Each anchor serves as a local viewpoint, focusing the model's attention on a specific region of the object's surface.
2. Spherical Distance Functions (SH)
From each anchor, the distance to the surface is encoded using Spherical Harmonics. This mathematical approach represents a function on a sphere using a small number of coefficients. By using SH, MASH achieves a very compact representation of local geometry, making it significantly easier for neural networks to process and learn.
3. The Mask (The "M" in MASH)
Standard spherical functions often struggle with discontinuities—areas where an object ends or where one part overlaps another. MASH solves this by introducing a generalized view cone with a free-form base. This mask precisely clips the spherical function to a specific local patch, allowing the model to represent sharp boundaries, holes, and complex overlaps without the artifacts typically associated with low-order spherical mappings.
To enhance accuracy, the authors also developed an Anchored Inverse Transformation. This geometric technique "bends" the surface around the anchor, making it easier for the spherical harmonics to fit the geometry accurately, even for complex or thin structures.
Optimization and Differentiability
MASH is designed to be fully differentiable. The authors developed an optimization pipeline that allows the representation to be "fitted" to raw, unoriented point clouds. By minimizing loss functions related to surface fitting, coverage, and boundary continuity between patches, the model can reconstruct high-fidelity surfaces from noisy or incomplete data. This local-first approach means the model does not need to learn a single global function for a complex object; it only needs to master simple local geometries, which is far more efficient.
Experimental Results
The researchers evaluated MASH using standard benchmarks, including ShapeNet-V2 for surface reconstruction and Objaverse-82K for diverse, large-scale 3D generation tasks.
Superior Reconstruction Fidelity
In comparative tests on ShapeNet, MASH demonstrated state-of-the-art performance in surface reconstruction. It achieved a Chamfer Distance of 4.944, significantly outperforming existing methods such as PGR (6.381) and ARO-Net (15.697). The representation is particularly effective at recovering thin structures—such as the legs of a table or thin wires—that other methods often fail to capture, instead merging them into "blobs."
Efficiency in Generation
Because the MASH representation is so compact, generative models trained on it converge much faster than those using traditional formats. The authors report that training and sampling times are less than one-third of those for competing methods like 3DShape2VecSet, even when using the same network backbone. This efficiency makes MASH a strong candidate for real-time 3D generation applications.
Multi-modal Versatility
MASH has proven successful in generating 3D shapes from a variety of inputs, including text prompts, single images, and partial point clouds. It maintains better topological consistency than current baselines like InstantMesh, particularly in complex multi-modal generation scenarios.
Hybrid Capabilities and Future Work
MASH acts as a hybrid bridge between implicit and explicit representations. It offers the smoothness of implicit functions alongside the local "editability" of explicit patches. This enables unique applications such as shape blending or part-conditioned completion, where specific anchors (e.g., the base of a chair) can be frozen while the rest of the shape is regenerated.
While highly effective, the current iteration of MASH focuses solely on geometry. Future research aims to incorporate textures and material properties into the representation. Additionally, the authors suggest that adaptive anchor placement—placing more anchors in areas of high geometric complexity—could further optimize the representation's efficiency. Scaling MASH to larger "foundation" models for 3D generation remains a primary goal for the next stage of development.
Summary updated