# Volumetric and Multi-View CNNs for Object Classification on 3D Data

**Authors:** Charles R. Qi, Hao Su, Matthias Niessner, Angela Dai, Mengyuan Yan, Leonidas J. Guibas  
**arXiv:** [1604.03265](https://arxiv.org/abs/1604.03265)  
**Date:** 2016-04-12  
**Summary page:** https://www.summarizepaper.com/en/paper/1604.03265v2/

## Key points

- 3D information is crucial for object classification
- Convolutional Neural Networks (CNNs) are the state-of-the-art methods used to address this problem
- There are two types of CNNs being developed: volumetric CNNs and multi-view CNNs
- Existing volumetric CNN architectures and approaches are unable to fully exploit the power of 3D representations, leading to a large gap in empirical results compared to multi-view CNNs
- Charles R. Qi, Hao Su, Matthias Niessner, Angela Dai, Mengyuan Yan, and Leonidas J. Guibas aim to improve both volumetric and multi-view CNNs by introducing two distinct network architectures of volumetric CNNs and examining multi-view CNNs with multi-resolution filtering in 3D
- Their proposed approaches outperform current state-of-the-art methods for both volumetric and multi-view CNNs
- Their research highlights the importance of improving existing approaches for object classification using 3D information
- The proposed network architectures can help fully exploit the power of 3D representations in both volumetric and multi-view settings, leading to more accurate object classification results

## Summary

The availability of 3D shape models and the ease of capturing them have made 3D information crucial for object classification. Convolutional Neural Networks (CNNs) are currently the state-of-the-art methods used to address this problem. However, there are two types of CNNs being developed: volumetric CNNs based on volumetric representations and multi-view CNNs based on multi-view representations. Empirical results from these two types of CNNs exhibit a large gap, indicating that existing volumetric CNN architectures and approaches are unable to fully exploit the power of 3D representations. In response, Charles R. Qi, Hao Su, Matthias Niessner, Angela Dai, Mengyuan Yan, and Leonidas J. Guibas aim to improve both volumetric CNNs and multi-view CNNs by introducing two distinct network architectures of volumetric CNNs and examining multi-view CNNs with multi-resolution filtering in 3D. Their goal is to provide a better understanding of the space of methods available for object classification on 3D data. Their study shows that their proposed approaches outperform current state-of-the-art methods for both volumetric CNNs and multi-view CNNs. They provide extensive experiments designed to evaluate underlying design choices, thus providing insights into the effectiveness of different methods for object classification on 3D data. Overall, their research highlights the importance of improving existing approaches for object classification using 3D information. The proposed network architectures can help fully exploit the power of 3D representations in both volumetric and multi-view settings, leading to more accurate object classification results.