PPC-MT: Parallel Point Cloud Completion with Mamba-Transformer Hybrid Architecture

Jie Li; Shengwei Tian; Long Yu; Xin Ning

cs.CVMar 1, 2026

PPC-MT: Parallel Point Cloud Completion with Mamba-Transformer Hybrid Architecture

Authors: Jie Li, Shengwei Tian, Long Yu, Xin Ning

Organizations: Xinjiang University, Urumqi, China · Institute of Semiconductors, Chinese Academy of Sciences, Beijing, China

Abstract

Existing point cloud completion methods struggle to balance high-quality reconstruction with computational efficiency. To address this, we propose PPC-MT, a novel parallel framework for point cloud completion leveraging a hybrid Mamba-Transformer architecture. Our approach introduces an innovative parallel completion strategy guided by Principal Component Analysis (PCA), which imposes a geometrically meaningful structure on unordered point clouds, transforming them into ordered sets and decomposing them into multiple subsets. These subsets are reconstructed in parallel using a multi-head reconstructor. This structured parallel synthesis paradigm significantly enhances the uniformity of point distribution and detail fidelity, while preserving computational efficiency. By integrating Mamba's linear complexity for efficient feature extraction during encoding with the Transformer's capability to model fine-grained multi-sequence relationships during decoding, PPC-MT effectively balances efficiency and reconstruction accuracy. Extensive quantitative and qualitative experiments on benchmark datasets, including PCN, ShapeNet-55/34, and KITTI, demonstrate that PPC-MT outperforms state-of-the-art methods across multiple metrics, validating the efficacy of our proposed framework.

Figures & tables

Fig. 1: Illustration of PCA-Guided Decomposition. Target point clouds are first sorted according to Principal Component Analysis (PCA) and then uniformly decomposed into several structured segments.

Fig. 2: The PPCMT framework consists of two primary components: (a) Predicted Point Cloud Generation and (b) Target Point Cloud Decomposition. The Predicted Point Cloud Generation component includes multiple modules, three of which are illustrated in this figure: (c) Feature Extractor, (d) Seed Generator, and (e) Multi-Head Reconstructor.

Fig. 3: The detailed structure of the Mamba Encoder.

Fig. 4: The detailed structure of the Transformer Decoder.

Method	CD- $\ell_{1}$ $\downarrow$	DCD $\downarrow$	EMD $\downarrow$	F1 $\uparrow$
FoldingNet [ 6 ]	14.31	-	-	-
TopNet [ 7 ]	12.15	-	-	-
AtlasNet [ 9 ]	10.85	-	-	0.616
CRN [ 12 ]	8.51	-	-	0.652
PCN [ 10 ]	9.64	-	-	0.695
GRNet [ 19 ]	8.83	0.622	-	0.708

TABLE I: Results on PCN dataset. (CD-

\ell_{1}\times 10^{3}

, EMD

\times 10^{3}

and F1-Score@1%)

Fig. 5: Qualitative comparison of PCN dataset. All methods above take the point clouds in the first column as inputs. We highlight some regions with red bounding box, which clearly show the effectiveness of our method.

Method	CD- $\ell_{1}$ $\downarrow$	CD- $\ell_{2}$ $\downarrow$	DCD $\downarrow$	EMD $\downarrow$	F1(mCls) $\uparrow$	F1 (Ova) $\uparrow$
FoldingNet [ 6 ]	-	3.12	0.082	-	-	0.082
TopNet [ 7 ]	-	2.91	0.126	-	-	0.126
PCN [ 10 ]	-	2.66	0.618	-	-	0.133
GRNet [ 19 ]	-	1.97	0.592	-	-	0.238
PoinTr [ 20 ]	13.95	1.05	0.539	38.86	0.468	0.499
SeedFormer [ 22 ]	13.97	0.92	0.552	37.09	0.444	0.473

TABLE II: Results on ShapeNet55. All reported results are averaged across three difficulty levels. (CD-

\ell_{1}\times 10^{3}

, CD-

\ell_{2}\times 10^{3}

, EMD

\times 10^{3}

and F1-Score@1%)

Method	34 seen categories					21 unseen categories
Method	CD- $\ell_{1}$ $\downarrow$	CD- $\ell_{2}$ $\downarrow$	DCD $\downarrow$	EMD $\downarrow$	F1 $\uparrow$	CD- $\ell_{1}$ $\downarrow$	CD- $\ell_{2}$ $\downarrow$	DCD $\downarrow$	EMD $\downarrow$	F1 (mCls) $\uparrow$	F1 (0va) $\uparrow$
FoldingNet [ 6 ]	-	2.35	-	-	0.139	-	3.62	-	-	-	0.095
TopNet [ 7 ]	-	2.31	-	-	0.171	-	3.50	-	-	-	0.121
PCN [ 10 ]	-	2.22	0.624	-	0.154	-	3.85	0.644	-	-	0.101
GRNet [ 19 ]	-	1.74	0.600	-	0.251	-	2.99	0.625	-	-	0.216
PoinTr [ 20 ]	15.07	1.09	0.604	42.35	0.402	17.64	1.75	0.629	46.30	0.370	0.360

TABLE III: Results on ShapeNet34. All reported results are averaged across three difficulty levels. (CD-

\ell_{1}\times 10^{3}

, CD-

\ell_{2}\times 10^{3}

, EMD

\times 10^{3}

and F1-Score@1%)

Fig. 6: Qualitative comparison of ShapeNet55. All methods above take the point clouds in the first column as inputs. We highlight some regions with red bounding box, which clearly show the effectiveness of our method.

Method	Fidelity $\downarrow$	MMD $\downarrow$	Consistency $\downarrow$	Uniformity $\downarrow$ for different $p$
Method	Fidelity $\downarrow$	MMD $\downarrow$	Consistency $\downarrow$	0.4%	0.6%	0.8%	1.0%	1.2%
GRNet [ 19 ]	0.816	0.568	0.407	2.69	4.98	7.42	9.93	12.46
PoinTr [ 20 ]	0.000	0.526	0.512	4.21	7.55	11.46	15.77	20.44
AdaPoinTr [ 23 ]	1.439	0.366	0.569	15.17	21.87	28.56	35.09	41.62
PPC-MT (Ours)	1.487	0.374	0.403	1.46	3.19	5.45	8.16	11.24

TABLE IV: Results on KITTI dataset. (Fidelity

\times 10^{3}

, MMD

\times 10^{3}

, Consistency

\times 10^{3}

)

Fig. 7: Qualitative comparison of KITTI. All methods above take the point clouds in the first column as inputs.

Rec. Heads	CD- $\ell_{1}$ $\downarrow$	DCD $\downarrow$	EMD $\downarrow$	F1 $\uparrow$
2	6.58	0.502	18.59	0.857
4 (ours)	6.60	0.491	17.43	0.860
8	6.71	0.494	17.09	0.855

TABLE V: Effect of different number of reconstruction heads. (CD-

\ell_{1}\times 10^{3}

, EMD

\times 10^{3}

and F1-Score@1%)

Method	CD- $\ell_{1}$ $\downarrow$	DCD $\downarrow$	EMD $\downarrow$	F-Score $\uparrow$
Random Dec.	6.61	0.495	17.57	0.858
Uniform Dec. (Ours)	6.60	0.491	17.43	0.860

TABLE VI: Effect of decomposition strategy. (CD-

\ell_{1}\times 10^{3}

, EMD

\times 10^{3}

and F1-Score@1%)

Method	Layers	Params	FLOPs	CD-L1 $\downarrow$	DCD $\downarrow$	EMD $\downarrow$	F1 $\uparrow$
Transformer Enc.	6	62.81M	33.48G	6.74	0.509	19.02	0.846
Transformer Enc.	12	75.43M	36.71G	6.63	0.494	17.69	0.857
Mamba Enc. (Ours)	12	55.50M	30.68G	6.60	0.491	17.43	0.860

TABLE VII: Effect of different architecture. (CD-

\ell_{1}\times 10^{3}

, EMD

\times 10^{3}

and F1-Score@1%)

Method	Params	FLOPs	CD-L1 $\downarrow$	DCD $\downarrow$	EMD $\downarrow$	F1 $\uparrow$
SeedFormer [ 22 ]	3.31M	53.76G	6.74	0.567	27.91	0.821
SVDFormer [ 25 ]	32.63M	39.26G	6.54	0.532	24.58	0.841
AdaPoinTr [ 23 ]	32.49M	15.08G	6.53	0.536	24.12	0.845
Ours	55.50M	30.68G	6.60	0.491	17.43	0.860

TABLE VIII: Complexity Analysis. (CD-

\ell_{1}\times 10^{3}

, EMD

\times 10^{3}

and F1-Score@1%)

Point Cloud Completion

Point cloud completion reconstructs missing 3D geometry from sparse, partial, noisy, or occluded point cloud observations.

See all 30 papers

I2PRef: Image-Driven Point Completion with Iterative Refinement
May 26, 2026Azhar Hussian, Marina Ritthaler, André Kaup +1Point Cloud Completion 3D Reconstruction
Friedrich-Alexander-Universität Erlangen-Nürnberg
SUMI: Scalable Unified Model for 3D Point Cloud Inference
Aug 8, 2026Yanlong LI, Kanchana ThilakarathnaPoint Cloud Completion 3D Shape Completion
MedPCFM: Improving Medical Point Cloud Completion by Integrating Point Transformers and Flow Matching
Jun 23, 2026Kamil Kwarciak, Marek WodzinskiFlow Matching Point Cloud Completion
Department of Measurement and Electronics, AGH University of Krakow, Krakow, Poland · Sano Centre for Computational Medicine, Krakow, Poland

May 26, 2026J/K move · Enter open · S save

I2PRef: Image-Driven Point Completion with Iterative Refinement

Azhar Hussian, Marina Ritthaler, André Kaup, Vasileios Belagiannis

Friedrich-Alexander-Universität Erlangen-Nürnberg

Point Cloud Completion 3D Reconstruction 3D Shape Completion Single-View 3D Reconstruction

Open paper AskPDF arXiv

We present an image-conditioned point cloud completion approach that treats images as the primary geometric source rather than a secondary guide. To this end, we introduce an Image-to-Point (I2P) module that can reconstruct complete point clouds directly from a single RGB image, with no need for 3D inputs. Additionally, we introduce a transformer-based Point-to-Point (P2P) refinement module that uses self- and cross-attention between point tokens and image features to iteratively refine the coarse I2P output. The I2P module enables the image encoder to learn rich geometric representations, while the P2P module progressively recovers fine-grained details. Unlike existing multimodal methods that rely on auxiliary losses or fusion modules, our explicit I2P task provides a strong, geometry-aware prior based on images alone. Extensive experiments on ShapeNet-ViPC demonstrate state-of-the-art completion performance with a 12.3% relative Chamfer Distance improvement over prior methods. Code is available at: https://github.com/AzharSindhi/I2PRef.git

Abstract

Point Cloud Completion

Explore similar work

I2PRef: Image-Driven Point Completion with Iterative Refinement

SUMI: Scalable Unified Model for 3D Point Cloud Inference

MedPCFM: Improving Medical Point Cloud Completion by Integrating Point Transformers and Flow Matching