# Voint Cloud: Multi-View Point Cloud Representation for 3D Understanding

**Authors:** Abdullah Hamdi, Silvio Giancola, Bernard Ghanem  
**arXiv:** [2111.15363](https://arxiv.org/abs/2111.15363)  
**Date:** 2021-11-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.15363v2/

## Key points

- Multi-view projection methods are effective for 3D understanding tasks such as classification and segmentation.
- The concept of multi-view point clouds (Voint cloud) has been introduced to combine multi-view methods with 3D point clouds.
- Voint cloud combines the compactness of 3D point clouds with the natural view-awareness of multi-view representations.
- VointNet, a neural network equipped with convolutional and pooling operations, can learn representations in the Voint space.
- VointNet achieves state-of-the-art performance on standard benchmarks for 3D classification, shape retrieval, and robust 3D part segmentation.
- The proposed approach outperforms existing methods under realistic rotated setups of ScanObjectNN and ShapeNet Parts.
- Compared to other widely used 3D representations, Voint cloud shares the view-dependency of NeRFs while inheriting the merits of explicit point clouds.
- This work presents a new approach for combining multi-view projection methods with widely available 3D point clouds through the use of Voint clouds which achieves state-of-the art results on various benchmarks for 3D understanding tasks.

## Summary

Multi-view projection methods have shown promising performance in 3D understanding tasks such as 3D classification and segmentation. To address the challenge of combining these multi-view methods with widely available 3D point clouds, researchers have introduced the concept of the multi-view point cloud (Voint cloud). This novel representation combines the compactness of 3D point clouds with the natural view-awareness of multi-view representations. The Voint cloud can be equipped with convolutional and pooling operations to enable a Voint neural network (VointNet) to learn representations in the Voint space. The proposed VointNet achieves state-of-the-art performance on standard benchmarks for 3D classification, shape retrieval, and robust 3D part segmentation (ScanObjectNN, ShapeNet Core55, and ShapeNet Parts). Furthermore, under realistic rotated setups of ScanObjectNN and ShapeNet Parts our approach outperforms existing methods. In comparison to other widely used 3D representations such as Multi-View Renderings, Voxels, Meshes, Explicit Point Clouds and Implicit Point Clouds; our proposed Voint cloud shares the view-dependency of NeRFs while inheriting the merits of explicit point clouds. Overall this work presents a new approach for combining multi-view projection methods with widely available 3D point clouds through the use of Voint clouds which achieves state-of-the art results on various benchmarks for 3D understanding tasks.