EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

CMPCC: Cross-Modal Point Cloud Completion Model Fusing Single-View Images and Partial Point Clouds

Liming Zhang, Chenxi Sun, Yuxing Shi, Tingcheng Zhang, Jinfeng Jia, Lu Gao

IEEE Access · 2026 · 14 · 82612-82636

DOI: 10.1109/access.2026.3696315 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Liming Zhang und 5 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Point cloud completion is a core task in 3D computer vision, aiming to recover the complete 3D shape of an object from partially incomplete point clouds collected by 3D scanning sensors. In practical scenarios, 3D scanning sensors are affected by factors such as occlusion, insufficient illumination, and resolution limitations, leading to frequent regional missing in the collected point clouds. However, most existing main-stream methods rely on unimodal shape priors and directly infer missing regions from local point clouds through encoder-decoder architectures. This easily results in the loss of local geometric details and inconsistency in global structure, making it difficult to meet the requirements of high-precision applications. To address the above issues, we reconstruct point cloud completion as a cross-modal collaboration task and propose a cross-modal point cloud completion model fusing single-view images and partial point clouds (called CMPCC). The model integrates complementary information from monocular images and local point clouds to achieve high-quality completion with global structure fidelity and refined local details. CMPCC consists of two stages: (1) the first stage is coarse-grained point cloud reconstruction, which uses CNN to extract global features from single-view images and generates a coarse-grained complete point cloud through deconvolution to ensure consistency between the overall contour and the real object. Then, a probability-based patch segmentation strategy is adopted to divide the coarse point cloud into several local patches, and patch-level features are extracted to lay the foundation for detail generation. (2) The second stage is dense point cloud generation, which uses DGCNN to extract fine-grained geometric features of local point clouds, capturing the surface texture and structural relationships of objects. Through a cross-attention fusion module, image features and point cloud features are fused to generate a global style code that constrains the consistency of the global structure. A multi-patch generator is used to fuse coarse-grained patch features with fine-grained geometric information, generating local patch point clouds with rich details. Finally, the patches are stitched together and sampled via FPS to obtain the complete point cloud. Experiments on the ShapeNet-ViPC and ModelNet-MPC datasets show that CMPCC significantly outperforms unimodal/cross-modal baseline methods: on ShapeNet-ViPC, the average F-Score reaches 0.869 (7.3% higher than XMFNet), and the average CD is $1.066\times 10^{-3}$ (22% lower than XMFNet); on ModelNet-MPC, CMPCC achieves an average CD of $1.267\times 10^{-3}$ and an average F-Score of 0.845, reducing the average CD by 16.4% and increasing the F-Score by 6.8% compared with XMFNet, and reducing the average CD by 9.0% and increasing the F-Score by 5.8% compared with EGIInet, demonstrating strong cross-dataset generalization and performance superiority of the model.

Zitieren

APA

Zhang, L., Sun, C., Shi, Y., Zhang, T., Jia, J., & Gao, L. (2026). CMPCC: Cross-Modal Point Cloud Completion Model Fusing Single-View Images and Partial Point Clouds. IEEE Access, 14, 82612-82636. https://doi.org/10.1109/access.2026.3696315

Vancouver

Zhang L, Sun C, Shi Y, Zhang T, Jia J, Gao L. CMPCC: Cross-Modal Point Cloud Completion Model Fusing Single-View Images and Partial Point Clouds. IEEE Access. 2026;14:82612-82636. doi:10.1109/access.2026.3696315

BibTeX
@article{Zhang2026,
  author = {Zhang, Liming and Sun, Chenxi and Shi, Yuxing and Zhang, Tingcheng and Jia, Jinfeng and Gao, Lu},
  title = {CMPCC: Cross-Modal Point Cloud Completion Model Fusing Single-View Images and Partial Point Clouds},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {82612--82636},
  doi = {10.1109/access.2026.3696315}
}
Kostenlos starten →