When a LiDAR scanner captures a chair, it rarely sees the whole object. Occlusions and limited viewing angles result in "partial" point clouds—sparse collections of coordinates that leave the back legs or the underside of the seat to the imagination. The goal of 3D shape completion is to fill in these blanks.
The fundamental challenge in this field is a tension between fidelity and hallucination. We want the model to be strictly faithful to the points we actually saw, but we also want it to "hallucinate" plausible structures where data is missing. Traditional reconstruction losses often fail here because they are symmetric: they penalize the model both for missing observed points and for adding points that weren't in the original (partial) scan.
The Problem with Symmetry
In standard point cloud tasks, researchers often use the Chamfer Distance (CD) or Earth Mover’s Distance (EMD). These metrics are symmetric; they measure how far the predicted points are from the ground truth and how far the ground truth points are from the prediction.
As Baoquan Chen and Xuelin Chen observe in their work on unpaired point cloud completion, symmetric metrics are counter-productive when the input is incomplete. If you force a complete 3D shape to match a partial scan using a symmetric loss, the model will try to "collapse" the entire object into the footprint of the partial data, effectively deleting the very parts it is supposed to reconstruct.
Unidirectional Distance as a Geometric Anchor
To solve this, a recurring strategy across recent research is the use of unidirectional (or one-way) distance metrics, such as the unidirectional Hausdorff distance or directional Chamfer distance.
The intuition is simple: we treat the partial input as a strict subset of the desired output. Instead of asking "Are these two shapes identical?", we ask "Does the completed shape contain all the points from the partial scan?"
- Faithfulness: The loss penalizes the model if any point from the original scan is far from the generated surface. This ensures the "identity" of the specific object is preserved.
- Slack for Hallucination: Crucially, the loss does not penalize the model for generating points in the missing regions. Because the metric is one-way, the model is free to synthesize new geometry—like the fourth leg of a chair—without being "pulled back" by the empty space in the partial input.
Rundi Wu and Xuelin Chen demonstrated that this constraint allows Generative Adversarial Networks (GANs) to produce multiple diverse completions for a single input. By combining this one-way loss with a style-modulated latent space, Wesley Khademi and Fuxin Li showed that models can capture the inherent uncertainty of missing geometry—generating different plausible backs for the same chair—while keeping the observed front perfectly intact.
From Unpaired Learning to Zero-Shot Completion
This "subset" logic is particularly powerful when we don't have paired training data (i.e., we have partial scans and complete shapes, but they aren't of the same objects). Xin Wen used a directional matching loss to anchor partial inputs within a cycle-consistency framework, allowing models to learn the "manifold" of complete shapes without ever seeing a ground-truth completion. Similarly, Zhen Cao and colleagues utilized unidirectional metrics to transfer knowledge from clean synthetic data to messy real-world scans.
The most extreme version of this idea appears in "test-time" or "zero-shot" frameworks like ComPC. In these systems, Tianxin Huang and Gim Hee Lee apply a preservation constraint that literally freezes the original points during optimization. While a 2D diffusion model "hallucinates" the missing parts from different viewpoints, the preservation constraint ensures the final 3D object maintains the high fidelity of the ground-truth sensor data.
The Trade-off: Spurious Surfaces
While unidirectional metrics provide the freedom to hallucinate, they can be too permissive. As Linus Härenstam-Nielsen and Daniel Cremers point out in their work on DiffCD, one-sided distances can lead to "spurious surfaces"—extra geometric structures that aren't supported by any data but aren't penalized by the loss. Their research suggests that while asymmetry is necessary for completion, finding the right balance (or "symmetric-ish" differentiable distances) is key to preventing the model from hallucinating noise into empty space.