MTCLnet: A Computationally Efficient and Adaptive Information Fusion Network for Cross-Modal Point Cloud Completion
- 24 February 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (7) , 10768-10777
Abstract
Attention-based multi-modal fusion frameworks have significantly advanced View-guided Point Cloud Completion (ViPC), which aims to recover a complete 3D point cloud from a partial one with the aid of a single-view image. However, existing methods commonly employ asymmetric cross-attention for point cloud and image feature fusion, in which one modality is designated as the query and the other as key-value pairs, resulting in an information flow biased toward the querying modality rather than enabling flexible feature extraction across modalities. Moreover, supplement completion loss that require the participation of all point cloud and image tokens introduce substantial computational overhead. To address these limitations, we propose Memory Transfer with Contrastive Learning Network termed MTCLnet, a novel framework that leverages contrastive learning to effectively align the feature spaces between point cloud and image modalities. Furthermore, we design a cross-attention-based module equipped with learnable tokens to adaptively fuse informative features across modalities. The integrated design enables more accurate and computationally efficient point cloud completion. Extensive experiments on the ShapeNet-ViPC dataset demonstrate that our method outperforms existing approaches, achieving both improved reconstruction quality and reduced computational complexity.Keywords
This publication has 43 references indexed in Scilit:
- Explicitly Guided Information Interaction Network for Cross-Modal Point Cloud Completion2024
- SeedFormer: Patch Seeds Based Point Cloud Completion with Upsample Transformer2022
- GRNet: Gridding Residual Network for Dense Point Cloud Completion2020
- TopNet: Structural Point Cloud Decoder2019
- PCN: Point Completion Network2018
- A Papier-Mache Approach to Learning 3D Surface Generation2018
- FoldingNet: Point Cloud Auto-Encoder via Deep Grid Deformation2018
- Shape Completion Using 3D-Encoder-Predictor CNNs and Shape Synthesis2017
- Deep Residual Learning for Image Recognition2016
- The farthest point strategy for progressive image samplingIEEE Transactions on Image Processing, 1997