Traditional approaches to 3D shape completion often treat the problem as a generative task: a neural network takes a partial point cloud, encodes it into a latent vector, and then "hallucinates" a complete set of points from scratch. While effective at capturing global structure, these generative models frequently suffer from "structural blurring," where fine-grained geometric details and the underlying topology of the original input are lost in the reconstruction process.
A growing body of research suggests a more precise alternative: treating shape completion as a deformation process. Instead of generating new points, these methods move existing points from an incomplete input (or a reference template) toward their target positions in the complete shape.
Moving Points Instead of Creating Them
The core intuition behind models like PMP-Net, developed by Xin Wen and Zhizhong Han, is that the partial input already contains high-quality geometric information that should be preserved. By framing completion as a deformation task, the network learns to rearrange points rather than inventing them.
This approach offers two primary advantages:
- Detail Preservation: Because the network moves the original points, the specific textures and local structures present in the visible regions are maintained rather than being replaced by noisy, generated approximations.
- Topology Consistency: Moving points along continuous paths helps preserve the manifold structure of the object. Generative models often produce "clumpy" or disconnected point sets because they lack an inherent mechanism to ensure that the generated points form a coherent surface.
The Power of Multi-Step Paths
Moving a point from a partial scan to its final position in a complete shape is a complex mapping. Attempting to do this in a single step often leads to large, erroneous displacements. To solve this, Xin Wen and Xiang Peng introduced the concept of multi-step point moving paths.
In PMP-Net++, the deformation is broken down into an iterative, coarse-to-fine process. In early steps, points move in large increments to establish the global skeleton of the missing parts. In later steps, the search radius for movement shrinks, allowing the network to refine the local geometry. This path-based approach ensures that each point follows a trajectory that minimizes total distance while maximizing shape accuracy, effectively "flowing" the partial shape into its complete form.
Completion as Registration
The deformation framework naturally extends to scenarios where a reference model is available. Oshri Halimi and Leonidas Guibas demonstrated that for deformable objects like humans, completion can be achieved by warping a complete reference model to match a partial scan. This preserves the identity and fine details of the reference while adapting it to the observed pose.
This logic also applies to tele-registration—aligning two partial scans with little overlap. Zihao Yan and Daniel Cohen-Or proposed CTF-Net, which treats registration and completion as interdependent tasks. By "imagining" the missing geometry through completion, the model creates artificial overlap that makes registration possible. Conversely, the spatial reasoning required to align two parts provides a stronger training signal for understanding the object's global structure.
Similarly, JCRNet, developed by Xiang Li and Yi Fang, uses a shared latent space to jointly optimize for both tasks. This synergy ensures that the registration is guided by the "underlying structural essence" of the complete object rather than being limited by fragmented surface patches.
Why It Works: The Intuition of Flow
The success of these methods lies in their shift from hallucination to transformation. By learning the "flow" of points—whether from an incomplete input to a target or from a reference template to an observation—these models leverage the geometric continuity of the 3D world.
Instead of asking a network to remember what every possible chair looks like, we ask it to learn the rules of how a chair's surface can be stretched and filled. This makes the problem more constrained, more geometrically grounded, and ultimately more precise.