CelloFill: Benchmark for Watertight Hole Filling

• CelloFill is a synthetic benchmark that evaluates visually plausible hole filling in watertight mesh reconstruction under large missing regions.
• It employs controlled synthetic corruptions and standardized clay renders to assess completion quality using perceptual metrics like FID, LPIPS, and CLIP.
• The benchmark emphasizes achieving compact, single-shell outputs by discouraging artifacts such as double shells and leaky solids.

Searching arXiv for the referenced work and any directly relevant mentions of CelloFill. CelloFill is a synthetic benchmark for watertight remeshing and hole filling introduced alongside CelloCut, where it is used to stress-test reconstruction under large, scan-like missing regions while preserving the object’s overall topology (Yang et al., 18 May 2026). Unlike benchmarks centered on geometric correspondence under arbitrary real-world defects, CelloFill isolates the visual plausibility of hole completion on controlled synthetic corruptions. Its design targets a specific failure regime in watertight conversion: methods may appear surface-accurate locally yet remain volumetrically inconsistent globally, producing artifacts such as double shells or leaky solids that are unsuitable for downstream volumetric use.

1. Definition and benchmark role

CelloFill exists to probe the hard case of large, structured holes similar to scanning artifacts and to ask whether a method completes plausible, compact single-shell solids that look correct from multiple views (Yang et al., 18 May 2026). In the CelloCut framework, this role is complementary to that of CelloScan. The two benchmarks address different failure modes, adopt different ground-truth definitions, and use different evaluation metrics.

Aspect	CelloFill	CelloScan
Core purpose	Visual plausibility of hole completion on controlled synthetic corruptions	Geometry and topology recovery on challenging real meshes
Ground truth	Original watertight GSO mesh	“Outer-surface” ground truth built by virtual scanning
Evaluation	Perceptual image metrics on standardized clay renders	Chamfer, Hausdorff, normals, and F1 thresholds

This separation matters methodologically. CelloScan targets in-the-wild inputs with ambiguous interior–exterior structure and severe defects such as self-intersections and near-zero-thickness elements. CelloFill, by contrast, is synthetic, hole-centric, and explicitly render-based. A common misconception is to treat it as a generic watertightness benchmark; more precisely, it is a benchmark for visually plausible hole filling under large missing regions, not a direct surface-distance benchmark (Yang et al., 18 May 2026).

2. Construction, source data, and defect model

CelloFill is derived from watertight meshes taken from Google Scanned Objects (GSO), from which the authors algorithmically introduce 6–12 holes per mesh (Yang et al., 18 May 2026). The benchmark targets hole filling on meshes, specifically triangle meshes, rather than point clouds. The original mesh remains globally watertight and topologically consistent; the corrupted input is an open, non-watertight triangle mesh with large missing patches but no artificial single-layer or self-intersection artifacts beyond those induced by the hole deletion.

The corruption patterns are designed to mimic real scanning failures. The reported defect types are grazing-angle dropouts, elongated gaps from motion or anisotropic sampling, and stripe-like scan-line omissions. Inputs therefore contain open boundaries and missing patches, while overall topology is preserved aside from these openings. Manifold or non-manifold specifics beyond the holes are not emphasized for CelloFill, which distinguishes it from CelloScan, where broader topological pathology is central (Yang et al., 18 May 2026).

Several dataset-level properties are intentionally or effectively left unspecified in the paper. The main text does not report the number of models, category labels, per-category counts, or train/validation/test splits, and it does not provide statistics such as average hole area, percentage of single-layer structures, or explicit topological complexity measures. Scale, units, coordinate conventions, and mesh resolution are also not specified. This omission is significant for reproducibility and cross-paper comparability, but it is consistent with the benchmark’s use as an evaluation set rather than a training corpus (Yang et al., 18 May 2026).

3. Ground truth and render-based evaluation

For CelloFill, the ground truth is the original watertight GSO mesh (Yang et al., 18 May 2026). Because the benchmark evaluates visually plausible completion rather than one-to-one pixel or point matching, both the completed output and the ground truth are rendered under the same standardized setup before scoring. The protocol uses untextured “clay” renders from six canonical views, with identical lighting and material parameters, using the Trellis rendering pipeline.

The benchmark then computes three perceptual metrics in image space. FID is reported with lower being better and measures distributional similarity of render sets to the ground-truth set. LPIPS is also lower-is-better and captures perceptual distance through deep features. CLIP score is higher-is-better and is described as a semantic/image-text alignment proxy adapted to image similarity, reported as a percentage-like score (Yang et al., 18 May 2026).

Two exclusions are central to the protocol. First, pixel-wise PSNR and SSIM are intentionally excluded because they are too sensitive to small pose or surface deviations in large-hole settings. Second, geometric metrics are not used on CelloFill at all; those are reserved for CelloScan. This design choice should not be read as a rejection of geometric evaluation in general. Rather, it reflects the benchmark’s narrower objective: standardized assessment of perceptual plausibility when large missing regions make direct surface correspondence less aligned with the task definition (Yang et al., 18 May 2026).

4. Relation to CelloCut’s constructive volumetric formulation

CelloFill is tightly coupled to the central claim of CelloCut: watertight remeshing should be posed as a volumetric partitioning problem rather than a surface-level repair task (Yang et al., 18 May 2026). The benchmark’s hole patterns are explicitly chosen to expose the weaknesses of methods that emphasize local surface preservation without enforcing a globally consistent interior–exterior partition. In such settings, projection-based, UDF-dilation, or coarse-grid pipelines can leave large gaps unfilled or produce thin spikes, voxel-like structures, or pseudo-watertight multi-layer shells.

CelloCut addresses this by constructing an unconstrained Delaunay tetrahedralization $DT(V)$ of a simplified thickened proxy surface $S_{\mathrm{sim}}$ and assigning a binary label to each tetrahedron $c$:

$L(c) \in \{0,1\},$

where $0$ denotes interior and $1$ denotes exterior. Initialization is obtained from a UDF offset:

$\phi(x) = UDF_M(x) - \epsilon,$

with centroid-based labeling

$$L^*(c) = \begin{cases} 0 & \text{if } \phi(x_c) < 0,\ 1 & \text{otherwise.} \end{cases}$$

The feasible set is one-sided: proxy-supported interior cells are preserved, while exterior-initialized cells remain relabelable,

$$F = \{L \mid L(c)\in\{0,1\},\; L^*(c)=0 \Rightarrow L(c)=0,\; \forall c\}.$$

This is implemented by a unary term $V(c,l)=+\infty$ if $S_{\mathrm{sim}}$0 and $S_{\mathrm{sim}}$1, and $S_{\mathrm{sim}}$2 otherwise. The graph-cut objective is

$S_{\mathrm{sim}}$3

with area-weighted face penalties

$S_{\mathrm{sim}}$4

where

$S_{\mathrm{sim}}$5

The completion prior $S_{\mathrm{sim}}$6 makes unsupported new boundaries expensive and biases the solution toward compact, single-shell fills. Max-flow/min-cut returns the global optimum $S_{\mathrm{sim}}$7. In all experiments, the default parameters are $S_{\mathrm{sim}}$8 of the UDF grid spacing, $S_{\mathrm{sim}}$9, and UDF/SDF grids at $c$0 (Yang et al., 18 May 2026).

For CelloFill, the importance of this construction is practical rather than purely formal. Elongated or misaligned gaps encourage double-layered closures, while grazing-angle deletions encourage spiky or voxel-like artifacts. The fill-aware area penalties explicitly discourage unsupported interface creation, which in turn strongly suppresses pseudo-watertight artifacts such as double shells (Yang et al., 18 May 2026).

5. Baselines, quantitative results, and qualitative behavior

CelloFill is evaluated against MeshFix, ManifoldPlus, fTetWild, Dora, and Craftsman (Yang et al., 18 May 2026). VolumeMesher is excluded in this benchmark because of frequent manifoldness violations, specifically non-watertight outputs, in broader experiments. Metrics are computed only on valid watertight outputs.

The reported visual metrics are as follows. MeshFix attains FID 29.4505, LPIPS 0.0513, and CLIP 98.66, but it has an overwhelmingly high failure rate to produce watertight outputs, with the failure numbers not reported. ManifoldPlus obtains FID 69.7035, LPIPS 0.1223, and CLIP 93.61. fTetWild obtains FID 39.9217, LPIPS 0.0555, and CLIP 97.09. Dora obtains FID 37.1267, LPIPS 0.0645, and CLIP 96.99. Craftsman obtains FID 30.7946, LPIPS 0.0504, and CLIP 97.88. CelloCut obtains FID 20.1733, LPIPS 0.0465, and CLIP 98.42 (Yang et al., 18 May 2026).

Within that comparison, CelloCut achieves the best FID and the best LPIPS among methods that reliably return watertight outputs, while maintaining a strong CLIP score (Yang et al., 18 May 2026). The qualitative interpretation given in the paper is consistent with the benchmark design: CelloCut closes large, elongated gaps with minimal unsupported interface, yielding compact, single-shell solids that look correct from multiple viewpoints and avoid double shells or interior fragments that some baselines produce.

The benchmark also exposes parameter sensitivities. The overfilling versus underfilling trade-off is governed by $c$1: too small a value encourages overfilling, whereas too large a value can leave large gaps. The paper reports that $c$2 in $c$3 is a good range, with default value $c$4. The thickening parameter $c$5 resolves single-layer ambiguities and stabilizes labeling; too small a value leaves residual holes or thin collapses, while too large a value oversmooths detail. The default $c$6 gives the best geometric trade-off in broader experiments (Yang et al., 18 May 2026).

A further design choice affects interpretation. Because the unary constraint preserves all proxy-supported interior, highly ambiguous shapes can converge to conservative but not necessarily author-intended topology. The paper characterizes this as a choice that favors watertight consistency and compactness over exact surface replication (Yang et al., 18 May 2026).

6. Practical protocol, reproducibility, and limitations

The practical evaluation pipeline for CelloFill is fully mesh-based and deterministic (Yang et al., 18 May 2026). All methods are compared through OBJ outputs. For CelloCut itself, the reported high-level pipeline is: compute a UDF on a $c$7 grid; set $c$8; extract $c$9 with Marching Cubes; decimate the resulting proxy surface to $L(c) \in \{0,1\},$0 with default approximately 95% face removal; tetrahedralize the simplified proxy vertices with CGAL; initialize $L(c) \in \{0,1\},$1 by $L(c) \in \{0,1\},$2; build the graph with unary constraints and fill-aware face-area pairwise terms; solve via max-flow/min-cut using PyMaxflow; extract the interface; convert it to an SDF; and re-extract the zero level set by Marching Cubes (Yang et al., 18 May 2026).

For benchmark evaluation, both the output mesh and the ground-truth mesh are rendered from six canonical clay views using the same Trellis renderer, cameras, lighting, and material setup. LPIPS, FID, and CLIP are then computed with standard implementations. The paper states that the project page at https://rangeryx-66.github.io/CelloCut/ hosts code and resources, but the main text does not specify a direct dataset link, license, or file organization for CelloFill (Yang et al., 18 May 2026).

The benchmark has clear limitations. It focuses on synthetic hole patterns on otherwise watertight meshes and therefore does not span the full spectrum of real defects, such as pervasive self-intersections, disconnected internal clutter, or mixed solid–shell assemblies. Those defects are instead better represented in CelloScan. In addition, the omission of dataset size, category counts, splits, and hole-size distributions limits direct reproducibility and weakens standardized comparison across later work (Yang et al., 18 May 2026).

The extensions proposed in the paper are correspondingly targeted. These include enriching CelloFill with controlled single-layer regions, self-intersections, and mixed solid–shell configurations; reporting dataset-wide statistics such as hole-area distributions and topology measures; and adding task-aligned volumetric consistency checks, including double-shell suppression metrics, alongside the present perceptual suite. This suggests a broader interpretation of CelloFill as an intentionally narrow but diagnostically sharp benchmark: it does not aim to cover all non-watertight failure modes, but rather to isolate a particularly consequential one—visually plausible, compact, single-shell completion under large missing regions (Yang et al., 18 May 2026).