# Toward Realistic Single-View 3D Object Reconstruction with Unsupervised Learning from Multiple Images

**Authors:** Long-Nhat Ho, Anh Tuan Tran, Quynh Phung, Minh Hoai  
**arXiv:** [2109.02288](https://arxiv.org/abs/2109.02288)  
**Date:** 2021-09-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2109.02288v2/

## Key points

- Recovering 3D structure of an object from a single image is challenging due to its ill-posed nature.
- Using numerous photos of the same object category can help learn a strong 3D shape prior for the object.
- Wu et al. (2020) demonstrated success with unsupervised learning but their algorithm only works for symmetric objects.
- Long-Nhat Ho et al. present a novel unsupervised algorithm that can learn a 3D reconstruction network from a multi-image dataset without requiring symmetry in the objects.
- Their algorithm covers both symmetry-required and non-symmetry-required scenarios, making it more general than previous approaches.
- They employ a novel albedo loss that improves reconstructed details and realisticity.
- The authors demonstrate that their method surpasses previous work in both quality and robustness through experiments on datasets with various structures including single-view, multi-view, image-collection, and video sets.
- Their approach achieves realistic single-view 3D object reconstruction without needing explicit supervision or imposing any symmetry assumptions.
- This paper presents an important contribution towards improving the accuracy and applicability of unsupervised learning based methods for 3D object reconstruction from images.

## Summary

The task of recovering the 3D structure of an object from a single image is challenging due to its ill-posed nature. One approach to address this challenge is to use numerous photos of the same object category to learn a strong 3D shape prior for the object. Wu et al. (2020) demonstrated the success of this approach by obtaining impressive 3D reconstruction networks with unsupervised learning. However, their algorithm only works for symmetric objects, limiting its applicability. In this paper, Long-Nhat Ho et al. present a novel unsupervised algorithm that can learn a 3D reconstruction network from a multi-image dataset without requiring symmetry in the objects. Their algorithm covers both symmetry-required and non-symmetry-required scenarios, making it more general than previous approaches. Additionally, they employ a novel albedo loss that improves reconstructed details and realisticity. The authors demonstrate that their method surpasses previous work in both quality and robustness through experiments on datasets with various structures, including single-view, multi-view, image-collection, and video sets. Their approach achieves realistic single-view 3D object reconstruction without needing explicit supervision or imposing any symmetry assumptions. Overall, this paper presents an important contribution towards improving the accuracy and applicability of unsupervised learning based methods for 3D object reconstruction from images.