DMF-Net: Image-Guided Point Cloud Completion with Dual-Channel Modality Fusion and Shape-Aware Upsampling Transformer

TL;DR

Extensive quantitative and qualitative experimental results show that DMF-Net outperforms the state-of-the-art unimodal and multimodal point cloud completion works on ShapeNet-ViPC dataset.

Abstract

In this paper we study the task of a single-view image-guided point cloud completion. Existing methods have got promising results by fusing the information of image into point cloud explicitly or implicitly. However, given that the image has global shape information and the partial point cloud has rich local details, We believe that both modalities need to be given equal attention when performing modality fusion. To this end, we propose a novel dual-channel modality fusion network for image-guided point cloud completion(named DMF-Net), in a coarse-to-fine manner. In the first stage, DMF-Net takes a partial point cloud and corresponding image as input to recover a coarse point cloud. In the second stage, the coarse point cloud will be upsampled twice with shape-aware upsampling transformer to get the dense and complete point cloud. Extensive quantitative and qualitative experimental results show that DMF-Net outperforms the state-of-the-art unimodal and multimodal point cloud completion works on ShapeNet-ViPC dataset.

Figures (7)

• Figure 1: The architecture of our proposed DMF-Net. It takes two stages to recover the complete point cloud. In the first stage, a coarse point cloud is generated according to the partial point cloud and corresponding single-view image. In the second stage, the coarse point cloud will be upsampled to get the dense and complete output.
• Figure 2: The architecture of 3D encoder, which is employed to extract the feature of partial point cloud.
• Figure 3: The architecture of Dual-channel Modality Fusion (DMF) module, which fuses the point-wise feature and the pixel-wise feature in a symmetric way.
• Figure 4: The architecture of decoder, which takes in the fused global feature and reconstructs a coarse point cloud.
• Figure 5: The architecture of upsampler, which consists of two consecutive shape-aware Upsample Transformers (SUTs) to generate the complete point cloud in a coarse-to-fine manner.