Corrupted Point Cloud Completion Dataset (CPCCD)

• CPCCD is an object-level benchmark that systematically introduces controlled corruptions in partial point clouds to evaluate model robustness.
• It retains paired complete ground truth from PCN while perturbing inputs with eight distinct corruption settings simulating sensor and occlusion errors.
• Empirical results reveal that transformer-based models outperform older methods, emphasizing the need for corruption-aware fine-tuning in completion tasks.

Searching arXiv for the cited papers to ground the article in current research. The Corrupted Point Cloud Completion Dataset (CPCCD) is an object-level benchmark for point cloud completion under explicit input corruption, introduced to evaluate whether completion models remain effective when partial point clouds are contaminated by clutter, occlusion, sensor-like perturbations, motion artifacts, scale mismatch, and viewpoint variation rather than merely missing geometry (Tesema et al., 22 Jul 2025). It is constructed as a corruption-rich extension of the PCN benchmark, preserving paired complete ground truth while perturbing the partial inputs across eight corruption settings. In the literature on point cloud completion, CPCCD occupies a distinct position between clean synthetic benchmarks and real-world but weakly standardized datasets: it retains controlled supervision while shifting evaluation toward robustness, which the source paper presents as the central practical deficiency of existing completion protocols (Tesema et al., 22 Jul 2025).

1. Research context and motivation

Point cloud completion methods have commonly been developed and evaluated under an idealized assumption: the input is partial but otherwise clean. The source paper situates CPCCD against benchmarks such as PCN, MVP, Completion3D, ShapeNet-based completion splits, KITTI, ScanNet, and a “Projected ShapeNet” variant, arguing that these datasets either provide clean synthetic partial/complete pairs or real scans without standardized object-level completion ground truth and controlled corruption conditions (Tesema et al., 22 Jul 2025). In that framing, strong benchmark performance does not necessarily indicate robustness to real acquisition artifacts.

This motivation aligns with earlier work showing that practical point cloud completion often fails once outliers or scene clutter are introduced. The CS-Net paper argued that standard completion datasets are unrealistic because real scans include surrounding objects and background points and because segmentation is imperfect; it therefore introduced a dataset in which incomplete point clouds contain outliers from neighboring clutter (Ma et al., 2022). A related development in urban-scale completion, Building-PCC, likewise emphasized that synthetic CAD-derived benchmarks do not reflect the irregularities of real airborne LiDAR data, where incompleteness arises from occlusion, absorption, viewpoint dependence, and density limitations (Gao et al., 2024). CPCCD extends this trajectory by making corruption itself the organizing principle of the benchmark rather than treating it as an incidental nuisance.

A plausible implication is that CPCCD reframes completion robustness as a first-class evaluation target. Instead of asking only whether a model can infer missing parts from a clean partial observation, it asks whether the model can recover object geometry when the observation is simultaneously incomplete and corrupted.

2. Dataset origin and construction

CPCCD is constructed from the PCN dataset, which contains partial/complete pairs for 30,974 shapes drawn from 8 categories: Airplane, Cabinet, Car, Chair, Lamp, Sofa, Table, and Boat (Tesema et al., 22 Jul 2025). In PCN, complete shapes are sampled from ShapeNet CAD models, and partial shapes are generated by back-projecting depth images from 8 viewpoints. CPCCD preserves the PCN train, validation, and test splits and applies corruption operators to the partial point clouds while keeping the complete ground-truth clouds unchanged.

The source paper states that CPCCD is effectively “8 PCN datasets,” meaning that each original partial cloud is expanded into 8 corruption variants (Tesema et al., 22 Jul 2025). The appendix reports 28,974 training objects, 800 validation objects, and 1,200 test objects, for a total of 30,974 objects. For corrupted partial point clouds, the appendix reports 231,792 training partial point clouds; the validation and test totals are described as reflecting 8 corruption versions per input partial, although the manuscript notes slight table-formatting inconsistency for those counts (Tesema et al., 22 Jul 2025). This preserves the category and split structure of PCN while altering only the observed partial inputs.

The design choice is methodologically important. Because CPCCD inherits PCN’s paired supervision, it remains suitable for objective quantitative evaluation using standard completion metrics. At the same time, by perturbing the input rather than the target, it isolates robustness to corrupted observations rather than changing the completion objective itself (Tesema et al., 22 Jul 2025).

3. Corruption taxonomy

CPCCD divides corruptions into two families: external corruptions, which add points not belonging to the target object, and internal corruptions, which alter or remove points from the object itself (Tesema et al., 22 Jul 2025). This taxonomy is broader than the single-noise-model setup used in many prior completion benchmarks.

External corruptions

External corruptions model contamination from unrelated scene content.

Corruption	Description	Parameters
External Object Interference (E_OI)	Randomly inserts points from simple geometric shapes standing in for unrelated objects	$N_o \in \{1,2,3\}$, $N_s = 12$, $N_p \in \{1/16, 1/12, 1/8, 1/4\}$, $N_d \in [0.05, 0.2]$
Background Interference: Wall (BI_w)	Adds points corresponding to a vertical planar surface such as a wall or door behind the object	$N_d \in [0.01, 0.05]$
Background Interference: Floor (BI_f)	Adds floor points visible under the object while removing occluded floor points	No continuous parameter domain listed

For E_OI, the 12 simple shapes are circle, square, rectangle, triangle, ellipse, hexagon, diamond, parallelogram, cylinder, sphere, cube, and pyramid (Tesema et al., 22 Jul 2025). The intent is to simulate nearby clutter entering the object crop without fully occluding the target.

Internal corruptions

Internal corruptions act directly on the target geometry.

Corruption	Description	Parameters
Occlusion by Other Objects (O_BOO)	Removes object points to simulate occlusion by another object	$N_o \in \{1,2,3,4\}$, same 12 shapes, $N_p \in \{1/8, 1/7, 1/6, 1/5, 1/4, 1/3\}$ of $N_t$, $N_d \in [0.05, 0.2]$
Dynamic Jitter with Trajectory (D_JT)	Displaces a subset of points along a trajectory to mimic motion-induced jitter	$J_a \in [0.01, 0.05]$, $N_s = 12$0 meters
Triaxial Rotation (T_R)	Rotates the partial point cloud around each axis	$N_s = 12$1 each from 0 to 10 degrees
Isometric Scaling (I_S)	Scales the partial point cloud to simulate object-size mismatch or sensor scale differences	Scale factor $N_s = 12$2

The eighth corruption setting is Random Combined Corruption (R_CC), which randomly combines any 2 to 7 of the seven primitive corruptions: $N_s = 12$3

$N_s = 12$4

This combined setting is explicitly presented as the most realistic and most difficult regime because it models compound disturbances rather than isolated degradations (Tesema et al., 22 Jul 2025).

A notable property of CPCCD is that it does not use a small number of fixed severity bins. Instead, corruption parameters are randomized within specified ranges and spatial constraints. This suggests an emphasis on continuous variability rather than discretized robustness evaluation.

4. Relation to earlier corrupted or real-world completion datasets

CPCCD is not the first dataset to challenge the clean-input assumption, but it differs from earlier efforts in scope and construction. The CS-Net work introduced a dataset built from ScanNet v2, ShapeNet, and Scan2CAD to simulate incomplete point clouds containing outliers from surrounding clutter, with segmentation labels marking CAD-object points as 1 and surrounding points as 0 (Ma et al., 2022). Its corruption model was scene-realistic clutter contamination rather than a broad corruption taxonomy. The dataset contained 21,054 models from 10 categories—chair, table, trash bin, TV/monitor, cabinet, bookshelf, sofa, lamp, bed, and tub—with train, validation, and test splits of 16,858, 2,096, and 2,100, respectively (Ma et al., 2022).

By contrast, CPCCD is built from PCN rather than from aligned real-scene crops, and it systematically spans clutter, occlusion, motion, rotation, scaling, and their random combinations (Tesema et al., 22 Jul 2025). It is therefore more controlled and more taxonomically explicit, though the source paper also states that its corruptions are simulated rather than captured directly from fully annotated real-world scans (Tesema et al., 22 Jul 2025).

Building-PCC provides a different point of comparison. It is a real-world benchmark for urban building completion using airborne LiDAR partial observations from AHN3 and AHN4 paired with complete LoD2-derived ground truth (Gao et al., 2024). There, incompleteness arises naturally from occlusion, tree blockage, glass absorption, viewpoint dependence, and resolution limitations rather than from explicit synthetic perturbation. The benchmark comprises about 50,000 buildings from The Hague and Rotterdam, split into 30,000 training, 10,000 validation, and 10,000 test instances, with complete point clouds sampled to 16,384 points per building (Gao et al., 2024).

These comparisons clarify CPCCD’s niche. It is paired and completion-specific like synthetic object benchmarks, but corruption-aware in a way standard synthetic datasets are not; it is more standardized than many real-world datasets, but less physically grounded than datasets whose incompleteness emerges directly from sensing. A plausible implication is that CPCCD is best interpreted as a robustness benchmark rather than as a full substitute for real-scan evaluation.

5. Evaluation protocol and metrics

The CPCCD paper evaluates completion models in two stages (Tesema et al., 22 Jul 2025). The first is zero-shot robustness, in which models trained on clean PCN are tested directly on CPCCD corruptions. The second is fine-tuned robustness, in which models are fine-tuned on CPCCD, specifically on the $N_s = 12$5 category, and then re-evaluated. This protocol separates failure due to distribution shift from gains due to corruption-aware adaptation.

The benchmark includes representative architectures from several model families: FoldingNet, PCN, GRNet, PoinTr, SnowFlakeNet, AdaPoinTr, and DWCNet (Tesema et al., 22 Jul 2025). The paper characterizes these families as graph-based, point-based, convolution-based, and transformer-based. The inclusion of both legacy and transformer-era methods is intended to reveal whether architectural advances on clean benchmarks translate into robustness under corruption.

CPCCD uses Chamfer Distance, reported as both $N_s = 12$6 and $N_s = 12$7, F-score, and Fidelity (Tesema et al., 22 Jul 2025). Chamfer Distance is defined as

$N_s = 12$8

with lower values indicating better geometric agreement. F-score is defined from point-wise precision and recall at threshold $N_s = 12$9: $N_p \in \{1/16, 1/12, 1/8, 1/4\}$0 where $N_p \in \{1/16, 1/12, 1/8, 1/4\}$1 and $N_p \in \{1/16, 1/12, 1/8, 1/4\}$2 denote precision and recall, respectively (Tesema et al., 22 Jul 2025). Fidelity measures how well the output preserves the input structure, and lower is better.

This metric set is notable because it does not evaluate only target-ground-truth proximity. The inclusion of Fidelity suggests an additional concern with structural consistency relative to the corrupted input, which is especially relevant when perturbations may induce the model either to overfit spurious observations or to ignore genuine structure.

6. Empirical findings and benchmark significance

The principal empirical conclusion is that current completion methods degrade substantially under corruption, even when they perform well on clean PCN (Tesema et al., 22 Jul 2025). Transformer-based models such as PoinTr, SnowFlakeNet, and AdaPoinTr are reported as strongest on the clean benchmark and generally stronger than older methods on corruptions as well, but CPCCD still exposes substantial failure cases. FoldingNet and PCN are much worse on clean and corrupted inputs than transformer methods, and GRNet is described as especially unstable on several corruptions and qualitatively among the worst methods (Tesema et al., 22 Jul 2025).

The paper identifies several corruption-specific trends. External Object Interference and Background Interference, both wall and floor, are the most damaging, while Random Combined Corruption is the hardest setting by a wide margin (Tesema et al., 22 Jul 2025). Isometric scaling and Triaxial rotation are also particularly challenging, even when local geometric detail is preserved. Occlusion by Other Objects and Dynamic Jitter are reported as somewhat less catastrophic for many models. The authors also note semantic misinterpretation under clutter: wall interference can make a chair look like a sofa, and floor interference can erase or distort chair legs (Tesema et al., 22 Jul 2025).

After fine-tuning on CPCCD, all methods improve on corrupted inputs, and the gap between methods narrows across corruption types (Tesema et al., 22 Jul 2025). DWCNet achieves the best or near-best performance across most settings, particularly on $N_p \in \{1/16, 1/12, 1/8, 1/4\}$3, but the paper emphasizes that these gains reflect improved tolerance rather than true removal of corruption. Even after fine-tuning, models often retain corrupted structures in their outputs; they adapt to corruptions rather than fully denoising them.

These results support several broader conclusions stated by the authors: current completion methods are not robust to realistic corruptions; external clutter is especially harmful; transformer-based completion models are generally more robust than point-based or convolution-based ones; point-based and graph-based methods can sometimes handle scale or rotation better but lose detail; fine-tuning on CPCCD improves tolerance but not true denoising; and the hardest corruption is random combined corruption $N_p \in \{1/16, 1/12, 1/8, 1/4\}$4 (Tesema et al., 22 Jul 2025). Taken together, these findings position CPCCD as a diagnostic benchmark for robustness failure modes rather than merely another leaderboard dataset.

7. DWCNet, limitations, and implications for future work

CPCCD is closely tied to the Denoising-While-Completing Network (DWCNet), which the source paper proposes as a completion framework with a Noise Management Module (NMM) (Tesema et al., 22 Jul 2025). The NMM uses a transformer-based clean branch and a multiscale convolution noisy branch, with contrastive losses that encourage clean features to align with clean ground truth while repelling noisy features. The paper reports that adding NMM improves performance both on PCN and especially on the corrupted $N_p \in \{1/16, 1/12, 1/8, 1/4\}$5 subset, reinforcing the claim that CPCCD captures a meaningful robustness problem (Tesema et al., 22 Jul 2025).

The benchmark’s limitations are explicitly stated. CPCCD’s corruptions are simulated rather than captured directly from fully annotated real-world scans; the simulation may not reproduce all sensor artifacts or scene dynamics; the dataset is limited to the 8 PCN object categories and indoor-object settings; and the authors plan to expand CPCCD to more categories, including ShapeNet34- and ShapeNet55-style categories, and to improve corruption realism (Tesema et al., 22 Jul 2025). These constraints matter for interpretation. CPCCD should not be treated as an exhaustive model of real-world scanning conditions.

At the same time, comparison with related datasets suggests why CPCCD remains methodologically useful. The CS-Net dataset shows that clutter-contaminated object completion already requires joint reasoning about segmentation and reconstruction (Ma et al., 2022). Building-PCC shows that real-world missingness is structurally imbalanced and can expose normalization mismatches, detail loss, and distributional gaps that synthetic benchmarks obscure (Gao et al., 2024). CPCCD contributes a controlled corruption space in which these robustness issues can be probed systematically at object level.

A plausible implication is that future work on point cloud completion will increasingly need to combine three benchmark properties rather than optimizing for only one: paired supervision, realistic corruption processes, and explicit robustness protocols. In that landscape, CPCCD functions as a bridge benchmark. It preserves the evaluation discipline of synthetic paired datasets while making corruption-rich robustness measurable in a standardized way (Tesema et al., 22 Jul 2025).