ProjFormer: Point Cloud Completion via Geometric-Projective Transformer and Cross-Modal Semantic Constraints

Sheng Liu; Meng Wang; Ruihui Li; Huilong Pi; Zhuo Tang; Kenli Li

cs.CVAug 15, 2026

ProjFormer: Point Cloud Completion via Geometric-Projective Transformer and Cross-Modal Semantic Constraints

Authors: Sheng Liu, Meng Wang, Ruihui Li, Huilong Pi, Zhuo Tang, Kenli Li

Organizations: Hunan University Chang Sha, China

Abstract

Point cloud completion is inherently ill-posed due to severe sparsity and ambiguity in partial observations. Existing multi-view methods alleviate this by incorporating 2D semantics, but often rely on learned attention and fixed fusion, which lack geometric consistency and adaptability. We propose ProjFormer, a cross-modal framework that enforces geometry-consistent 2D-3D interaction through explicit projection and adaptive feature routing. A Projective Guided View Attention module aligns 3D points with multi-view features via deterministic projection, enabling efficient and geometrically consistent aggregation. Building on this, a geometry-aware routing network performs point-wise adaptive fusion of structural and observation-driven features for progressive refinement. Experiments show that, under a lightweight design, ProjFormer delivers competitive performance with improved structural completeness.

Figures & tables

Figure 1. Unlike global 2D–3D fusion, ProjFormer deterministically aligns multi-view features with 3D points, enabling geometry-aware refinement with fewer artifacts and lower latency. Comparison between existing 2D–3D fusion methods and our ProjFormer. Existing methods use global pooling, while ours uses deterministic projection for fine-grained semantic injection.

Figure 2. Overview of ProjFormer. The partial point cloud is processed by parallel 3D and CCM-based multi-view branches. PGVA uses explicit projection to construct point-aligned view tokens, which are combined with seed points and the partial input through Geometry-Aware Routing Refinement to generate the completed point cloud.

Figure 3. Projective Guided View Attention. Each observed 3D point is projected onto multiple CCM views. View features and canonical coordinates are retrieved at valid projected locations and aggregated using geometry-aware, depth-gated weights to produce a point-aligned token.

Figure 4. Geometry-Aware Routing Refinement. Two cascaded refinement blocks adaptively balance structural features against an observation-conditioned fusion of local geometric and point-aligned view features to predict residual coordinate offsets.

Method	Venue	Average	Plane	Cabinet	Car	Chair	Lamp	Sofa	Table	Vessel
PCN	3DV’18	10.60	6.09	11.55	9.51	12.33	12.57	12.87	9.87	9.98
PoinTr	ICCV’21	8.38	4.75	10.47	8.68	9.39	7.75	10.93	7.78	7.29
SeedFormer	ECCV’22	6.74	3.85	9.05	8.06	7.06	5.21	8.85	6.05	5.85
SVDFormer	ICCV’23	6.55	3.64	8.81	7.44	6.93	5.32	8.60	5.86	5.80
CRA-PCN	AAAI’24	6.56	3.62	8.77	7.00	6.92	5.46	8.59	6.27	5.86
ODGNet	AAAI’24	6.50	3.77	8.77	7.56	6.84	5.09	8.47	5.84	5.66

Table 1. Quantitative comparison of point-cloud completion results on the PCN dataset measured by the

\mathrm{CD}_{\ell_{1}}

distance (

\times 10^{3}

Method	ShapeNet 55 dataset			34 seen categories			21 unseen categories
Method	CD-S ( $\downarrow$ )	CD-M ( $\downarrow$ )	CD-H ( $\downarrow$ )	CD-S ( $\downarrow$ )	CD-M ( $\downarrow$ )	CD-H ( $\downarrow$ )	CD-S ( $\downarrow$ )	CD-M ( $\downarrow$ )	CD-H ( $\downarrow$ )
SeedFormer	0.50	0.77	1.49	0.48	0.70	1.30	0.61	1.07	2.35
ODGNet	0.47	0.70	1.32	0.44	0.64	1.14	0.59	1.01	2.26
SVDFormer	0.48	0.70	1.30	0.46	0.64	1.13	0.61	1.05	2.19
CRA-PCN	0.48	0.71	1.37	0.45	0.65	1.18	0.55	0.97	2.19
GeoFormer	0.41	0.64	1.25	0.39	0.57	1.05	0.55	0.99	2.15

Table 2. Comparison with state-of-the-art methods on the ShapeNet-55/34 benchmarks using

\mathrm{CD}_{\ell_{2}}\times 10^{3}

. Lower is better (

\downarrow

Figure 5. Visual comparisons on the PCN dataset.

No.	Model Configuration	CCM	Point-aligned token	Depth Gating	PGVA	$\mathrm{CD}_{\ell_{1}}$ ( $\downarrow$ )	$\mathrm{CD}_{\ell_{2}}$ ( $\downarrow$ )	F-Score@1% (%) ( $\uparrow$ )
1	Full w/o CCM		✓	✓	✓	6.785	2.0466	82.1679
2	Full w/o Point-aligned	✓		✓	✓	6.448	1.8129	84.1216
3	Full w/o depth gating	✓	✓		✓	6.384	1.7599	84.4839
4	Full w/o PGVA	✓				6.493	1.8468	83.6134
5	Full Model	✓	✓	✓	✓	6.3446	1.7905	84.9282

Table 3. Ablation study on 4 different model components.

Figure 6. Qualitative completion results on the KITTI dataset.

Methods	PoinTr	SeedFormer	Ours
MMD $\downarrow$	8.21	1.09	0.48
FD $\downarrow$	0.00	1.45	1.97

Table 4. Quantitative comparison on KITTI

Method	Venue	$\mathrm{CD}_{\ell_{1}}$	Infer Time(ms)
CRA-PCN	AAAI’24	6.56	67.61
ODGNet	AAAI’24	6.50	45.39
SVDFormer	ICCV’23	6.55	69.40
FSC	CVPR’24	7.02	26.97
GeoFormer	ACMMM’24	6.42	26.37
Ours	ACMMM’26	6.34	15.85

Table 5. Comparison of completion accuracy and inference speed on the PCN dataset.

Point Cloud Completion

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

See all 30 papers

MAGE: View-guided Point Cloud Completion with Efficient Modality Alignment and Adaptive Geometry Enhancement
Jun 30, 2026Weize Quan, Zhengwei Wu, Kai Wang +1Point Cloud Completion Cross-Modal Alignment
MAIS, Institute of Automation, Chinese Academy of Sciences, Beijing, China. · School of Artificial Intelligence, University of Chinese Academy Sciences, Beijing, China. · GIPSA-lab, Univ. Grenoble Alpes, CNRS, Grenoble INP, 38000 Grenoble, France.
SplAttN: Bridging 2D and 3D with Gaussian Soft Splatting and Attention for Point Cloud Completion
May 2, 2026Zhaoyang Li, Zhichao You, Tianrui LiCross-Modal Learning Point Cloud Completion
School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China
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

Jun 30, 2026J/K move · Enter open · S save

MAGE: View-guided Point Cloud Completion with Efficient Modality Alignment and Adaptive Geometry Enhancement

Weize Quan, Zhengwei Wu, Kai Wang, Dong-Ming Yan

MAIS, Institute of Automation, Chinese Academy of Sciences, Beijing, China. · School of Artificial Intelligence, University of Chinese Academy Sciences, Beijing, China. · GIPSA-lab, Univ. Grenoble Alpes, CNRS, Grenoble INP, 38000 Grenoble, France.

Point Cloud Completion Cross-Modal Alignment Attention Mechanisms

Open paper AskPDF arXiv

View-based point cloud completion aims to recover a complete 3D shape from a partial point cloud, guided by a single-view image. However, existing approaches often suffer from limited performance due to weak modality alignment and limited self-geometry enhancement. To overcome these challenges, we propose a unified geometry-aware framework that integrates efficient modality alignment and adaptive geometry enhancement, mainly to address cross-modal geometric inconsistency of view-guided point cloud completion. Specifically, we propose a geometry-aware modality alignment by integrating a shared self-attention Transformer and cross-modality reconstruction supervision, which aims to bring features of the image and point cloud close to each other in a shared latent space describing the 3D object. To enhance the perception of global shape and local geometric details, we propose an adaptive geometry-aware self-attention module, which simultaneously considers local geometry-aware attention computation and the spatially-variant feature fusion. In addition, we apply a geometry-perceptive anchor refinement module to reorganize the anchor points (representing a local region of the shape) under appropriate supervision, further boosting the completion performance of our method. Extensive experiments on both synthetic and real-world datasets demonstrate that our method achieves superior performance over existing approaches. Our code will be available at https://github.com/weizequan/MAGE.

Abstract

Point Cloud Completion

Explore similar work

MAGE: View-guided Point Cloud Completion with Efficient Modality Alignment and Adaptive Geometry Enhancement

SplAttN: Bridging 2D and 3D with Gaussian Soft Splatting and Attention for Point Cloud Completion

I2PRef: Image-Driven Point Completion with Iterative Refinement