CelloScan: Benchmark for Watertight Remeshing

• CelloScan is a benchmark dataset that defines geometric fidelity by comparing reconstructed watertight solids to a virtual-scanned outer surface.
• It evaluates methods on raw, topologically challenging meshes with defects like open holes, self-intersections, and non-manifold structures.
• The benchmark uses metrics such as Chamfer Distance, Hausdorff Distance, ANC, and [email protected] to quantify the quality of watertight remeshing and solid reconstruction.

Searching arXiv for the specified paper to ground the article in the primary source. arXiv search query: ([2605.17853](/papers/2605.17853)) [CelloCut](https://www.emergentmind.com/topics/cellocut) CelloScan CelloScan is a benchmark dataset introduced in the watertight remeshing framework "CelloCut: Constructive Watertight Remeshing via Tetrahedral Cell Cuts" (Yang et al., 18 May 2026). In that context, it is a curated evaluation suite of real, defective, non-watertight triangle meshes designed to assess geometric fidelity on raw scans and scan-like assets under severe topological ambiguity. Unlike conventional repair benchmarks that compare outputs directly to defective input surfaces, CelloScan evaluates recovery of an object’s visible outer boundary via a virtual-scanning protocol, thereby aligning evaluation with the solid reconstruction objective rather than with the preservation of surface artifacts (Yang et al., 18 May 2026).

1. Definition and scope

In the CelloCut framework, CelloScan is explicitly a dataset / benchmark, not a component of the reconstruction algorithm itself (Yang et al., 18 May 2026). It was introduced together with CelloFill to support what the authors describe as a broader attempt to standardize evaluation for watertight remeshing. Its designated role is geometric fidelity on raw scans, whereas CelloFill emphasizes visual quality in hole filling and ModelNet10 is used for robustness testing under relatively benign conditions (Yang et al., 18 May 2026).

The benchmark targets inputs that are problematic for standard input-reconstruction protocols. These meshes exhibit severe non-watertightness, missing regions, self-intersections, and complex topology, so they do not define a unique interior volume in the usual sense (Yang et al., 18 May 2026). This is central to the rationale for CelloScan: a repaired output can be superficially close to the input surface while still being volumetrically inconsistent, for example by containing closely spaced double shells or pseudo-watertight structures (Yang et al., 18 May 2026).

A common misconception is to treat CelloScan as if it were an algorithm, a scanning system, or a learned model. In the CelloCut paper it is none of these. It is a benchmark specifically constructed to stress-test watertight remeshing and solid reconstruction methods on raw, topologically challenging meshes (Yang et al., 18 May 2026).

2. Construction and data characteristics

CelloScan is constructed from topologically challenging 3D meshes collected from Objaverse, the Tencent Hunyuan3D Watertight Conversion Challenge, and additional in-the-wild sources (Yang et al., 18 May 2026). The inputs are triangle meshes resembling raw scans or raw assets from content libraries and generative pipelines rather than synthetic degradations of clean CAD models. This origin matters because the benchmark is intended to reflect real defect patterns rather than controlled perturbations.

The defect profile is deliberately severe. The benchmark includes open holes, thin single-layer structures, self-intersections, and mixed combinations of these defects (Yang et al., 18 May 2026). Such pathologies undermine any unique inside-outside interpretation of the mesh and make purely local, surface-based repair intrinsically ambiguous. In the CelloCut paper, these are precisely the cases that expose failures such as double shells, leaky solids, and non-manifold or internally fragmented reconstructions (Yang et al., 18 May 2026).

The paper does not specify the number of shapes, category distribution, or train/validation/test splits for CelloScan (Yang et al., 18 May 2026). This omission is consistent with its use as an evaluation benchmark rather than as a learning dataset. The text reports average-per-model statistics over the benchmark, which suggests a non-trivial corpus, but any stronger quantitative statement about scale would be speculative.

3. Task formulation and relation to watertight remeshing

The task posed by CelloScan is to convert a defective input mesh into a strictly watertight, vertex-manifold, single-shell solid with a globally consistent interior-exterior partition, while remaining geometrically faithful to the object’s visible outer surface (Yang et al., 18 May 2026). The benchmark is therefore aligned with the volumetric interpretation of watertight conversion advocated by CelloCut.

In the primary formulation used on CelloScan, space is partitioned by Delaunay tetrahedralization of a thickened proxy surface, yielding tetrahedral cells $c \in T$ (Yang et al., 18 May 2026). Each tetrahedron is assigned a binary label

$$L(c) \in \{0, 1\},\quad 0 = \text{interior},\ 1 = \text{exterior}.$$

The feasible set preserves one-sided interior evidence from the proxy initialization $L^*$: $$F = \{ L \mid L(c)\in\{0,1\},\ L^*(c) = 0 \Rightarrow L(c)=0,\ \forall c\}.$$ Optimization is performed through graph-cut energy minimization: $L_{\text{opt}} = \arg\min_{L\in F} E(L),$ with energy

$$E(L) = \sum_c V(c, L(c)) + \sum_{(c_i,c_j)\in N} D(c_i, c_j)\,\mathbf{1}[L(c_i) \neq L(c_j)].$$

The unary term enforces one-sided interior anchoring,

$$V(c,l) = \begin{cases} +\infty, & L^*(c)=0 \ \text{and}\ l=1,\ 0, & \text{otherwise}, \end{cases}$$

and the pairwise term uses fill-aware interface regularization,

$D(c_i,c_j) = W_{ij}\,\mathrm{area}(f_{ij}),$

with

$$W_{ij} = \begin{cases} \lambda_{\text{fill}}, & L^*(c_i) = L^*(c_j),\ 1, & L^*(c_i) \neq L^*(c_j). \end{cases}$$

The default hyperparameter is $\lambda_{\text{fill}} = 20$ (Yang et al., 18 May 2026).

CelloScan does not alter this formulation. What distinguishes it is the difficulty of its inputs and the fact that evaluation is decoupled from the defective raw mesh. This suggests that CelloScan is designed less as a benchmark for local mesh cleanup than as a benchmark for globally consistent solid reconstruction.

4. Ground truth design and evaluation protocol

The defining methodological feature of CelloScan is that the raw input mesh is not treated as ground truth (Yang et al., 18 May 2026). Instead, the benchmark constructs a reference outer surface through a virtual scanning (ray-casting) procedure. By aggregating externally observed geometry from multiple viewpoints, this protocol approximates the object’s visible outer surface while suppressing interior clutter and defect-specific artifacts (Yang et al., 18 May 2026).

This procedure produces an Outer Surface GT point cloud against which reconstructed outputs are compared (Yang et al., 18 May 2026). The choice is task-aligned: it rewards recovery of the effective enclosing boundary rather than fidelity to broken, topologically ambiguous, or internally inconsistent input geometry. The authors emphasize that this protocol does not encode any bias toward CelloCut’s volumetric formulation and does not assume a particular reconstruction strategy (Yang et al., 18 May 2026).

The metrics reported on CelloScan are geometric metrics between sampled output surfaces and the Outer Surface GT point cloud (Yang et al., 18 May 2026). They are listed below.

Metric	Role in evaluation	Direction
Chamfer Distance (CD)	Surface distance to Outer Surface GT	Lower is better
Hausdorff Distance (HD)	Worst-case deviation	Lower is better
Absolute Normal Consistency (ANC)	Surface normal alignment	Higher is better
F-score ([email protected])	Precision/recall at fixed distance threshold	Higher is better

The paper identifies F1 as being computed under a fixed distance threshold, specifically [email protected] on CelloScan (Yang et al., 18 May 2026). Chamfer Distance is described as Chamfer-L2 distance, and ANC is bounded above by 1.0 in the usual interpretation (Yang et al., 18 May 2026). These choices collectively emphasize outer-surface fidelity rather than conformity to pathological input topology.

5. Quantitative and qualitative benchmark results

The CelloCut paper evaluates MeshFix, ManifoldPlus, fTetWild, Dora, Craftsman (CraftsMan3D), and CelloCut on CelloScan (Yang et al., 18 May 2026). The table in the provided text is partially affected by formatting noise, but the paper’s narrative interpretation is explicit: CelloCut achieves the best overall results in CD, ANC, and [email protected], while ManifoldPlus and Dora obtain slightly lower HD yet underperform on CD and F1, indicating sensitivity to local inconsistencies (Yang et al., 18 May 2026).

The reported values preserved in the source material are as follows.

Method	Failure / CD / HD / ANC / [email protected]	Reported outcome
MeshFix	0.202739 / 0.902212 / 0.5638 / 3.48 / unclear in snippet	Clearly worse overall
ManifoldPlus	0.000074 / 0.083855 / 0.9363 / 95.65 / unclear in snippet	Strong baseline
fTetWild	0.000808 / 0.134098 / 0.9057 / 84.38 / unclear in snippet	Weaker on all metrics
Dora	0.000623 / 0.084281 / 0.9402 / 92.94 / unclear in snippet	Competitive but inconsistent
Craftsman	0.000463 / 0.103374 / 0.9327 / 90.48 / unclear in snippet	Inferior to leading methods
CelloCut	0.000048 / 0.089819 / 0.9452 / 96.96 / best overall trend	Best CD, ANC, [email protected]

The accompanying textual summary states that CelloCut achieves the lowest CD, highest ANC, and highest [email protected] on CelloScan, with HD slightly higher than that of ManifoldPlus and Dora (Yang et al., 18 May 2026). This pattern is interpreted in the paper as evidence that CelloCut more accurately recovers visible boundaries while avoiding the local inconsistencies that degrade CD and F1 in competing methods.

Qualitative examples on CelloScan-type inputs include bad_doll and dinosaur (Yang et al., 18 May 2026). These cases show large semantic openings, articulated gaps, self-intersections, non-manifold structures, and other scan artifacts. On bad_doll, the paper reports that MeshFix retains only a small fragment, ManifoldPlus introduces sharp spikes, Dora forms a double-shell structure, fTetWild fails to close larger openings, and Craftsman exhibits voxel-like artifacts, whereas CelloCut produces a clean, single-shell watertight surface without internal fragments (Yang et al., 18 May 2026). On dinosaur, the paper describes CelloCut as robustly resolving holes and self-intersections while producing a smooth watertight mesh with a consistent interior-exterior definition suitable for downstream tasks (Yang et al., 18 May 2026).

These examples clarify the intended stress profile of CelloScan. The benchmark is not only measuring geometric distance; it is designed to reveal pseudo-watertight artifacts, non-manifold interiors, hollow double shells, and globally inconsistent solids that simpler protocols may overlook.

6. Position within the benchmark suite and implementation-specific analyses

CelloScan is one of three complementary evaluation settings in the CelloCut paper (Yang et al., 18 May 2026). ModelNet10 assesses robustness and topological validity on relatively clean CAD models. CelloFill is a synthetic benchmark built from Google Scanned Objects with algorithmically introduced holes and is evaluated largely through perceptual image metrics such as LPIPS, FID, and CLIP score on rendered clay images (Yang et al., 18 May 2026). CelloScan, by contrast, focuses on geometric fidelity to the visible outer boundary under severe real-world topological defects (Yang et al., 18 May 2026).

This division of labor is important. A method can perform well on nearly watertight CAD models or on visually plausible hole filling yet still fail on raw meshes with topological ambiguity. CelloScan occupies that hardest regime. A plausible implication is that it is intended as the paper’s most stringent test of whether a reconstruction method can produce compact, single-shell, volumetrically consistent solids under realistic failure modes.

The paper also reports parameter studies and runtime analyses specifically measured on CelloScan (Yang et al., 18 May 2026). The unsigned distance field used for the thickened proxy is computed on a $$L(c) \in \{0, 1\},\quad 0 = \text{interior},\ 1 = \text{exterior}.$$0 grid, with thickening offset $$L(c) \in \{0, 1\},\quad 0 = \text{interior},\ 1 = \text{exterior}.$$1 of the bounding box (Yang et al., 18 May 2026). Table 4, as described in the supplied details, shows that this setting gives the best balance on CelloScan: smaller $$L(c) \in \{0, 1\},\quad 0 = \text{interior},\ 1 = \text{exterior}.$$2 leaves gaps and fragments, whereas larger $$L(c) \in \{0, 1\},\quad 0 = \text{interior},\ 1 = \text{exterior}.$$3 oversmooths details and reduces [email protected] (Yang et al., 18 May 2026).

For the fill penalty, the authors evaluate

$$L(c) \in \{0, 1\},\quad 0 = \text{interior},\ 1 = \text{exterior}.$$4

on CelloScan while keeping all other settings fixed (Yang et al., 18 May 2026). Small values make new boundaries cheap and can induce overfilling; very large values adhere too strongly to the thickened proxy and leave large holes unresolved; moderate values, especially 10–20, yield the best trade-off, and 20 is chosen as the default (Yang et al., 18 May 2026).

The supplementary runtime figures identify CelloScan as a computationally demanding benchmark. Average runtime per model on CelloScan is reported as approximately 74.6 s per shape for CelloCut, compared with 19.6 s for Dora, 110.3 s for Craftsman, and 11.0 s for ManifoldPlus (Yang et al., 18 May 2026). For some methods, runtimes are not reported because of frequent failures or timeouts on challenging inputs. Within CelloCut, the runtime breakdown on CelloScan is dominated by graph-cut optimization (64.6%), followed by initial labeling + adjacency construction (20.1%), final extraction via SDF + Marching Cubes (10.2%), tetrahedralization (3.4%), and mesh decimation (1.5%), with UDF computation and initial Marching Cubes contributing about 0.1% each (Yang et al., 18 May 2026). This supports the paper’s interpretation that strong volumetric guarantees come at the cost of global optimization.

7. Significance, interpretation, and limitations

CelloScan is significant because it redefines what counts as successful evaluation for watertight remeshing on defective real-world meshes (Yang et al., 18 May 2026). Instead of asking whether an output closely matches a broken input surface, it asks whether the method recovers a valid enclosing boundary that agrees with the object’s externally visible geometry. This is a methodological shift from surface repair toward solid reconstruction.

The benchmark also functions as a diagnostic instrument for algorithmic failure modes. It exposes cases where a method is watertight only in a superficial sense, where a solid contains double shells, internal fragments, leaks, or non-manifold regions despite appearing locally plausible (Yang et al., 18 May 2026). By construction, it penalizes exactly those artifacts that arise when inside-outside ambiguity is resolved only locally.

Several limitations remain explicit. The paper does not provide dataset cardinality, category-level statistics, or split definitions for CelloScan (Yang et al., 18 May 2026). The project page is given, and the framing as a newly introduced benchmark strongly suggests intended public availability, but the supplied source material does not explicitly state release terms or licensing (Yang et al., 18 May 2026). Any more precise claim about dataset governance would therefore exceed the evidence.

Within the literature represented here, CelloScan should be understood narrowly and precisely: a newly curated benchmark for geometric fidelity on raw, topologically challenging meshes in watertight remeshing and solid reconstruction, introduced by CelloCut and evaluated against an outer-surface ground truth derived by virtual scanning (Yang et al., 18 May 2026). Its primary contribution is not a new reconstruction operator but a task-aligned standard for measuring whether reconstructed solids are both geometrically faithful and volumetrically coherent under extreme mesh defects.