# Unsupervised 3D Shape Completion through GAN Inversion

**Authors:** Junzhe Zhang, Xinyi Chen, Zhongang Cai, Liang Pan, Haiyu Zhao, Shuai Yi, Chai Kiat Yeo, Bo Dai, Chen Change Loy  
**arXiv:** [2104.13366](https://arxiv.org/abs/2104.13366)  
**Date:** 2021-04-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.13366v2/

## Key points

- ShapeInversion is a novel method for 3D shape completion using GAN inversion.
- It does not require paired training data and instead uses a pre-trained GAN to reconstruct partial inputs into complete shapes.
- ShapeInversion outperforms unsupervised methods and is comparable to supervised methods trained with paired data.
- It can handle real-world scans and various forms of partial inputs.
- ShapeInversion can produce multiple valid complete shapes for ambiguous partial inputs, as well as enable shape manipulation and interpolation.
- The authors provide additional details about their approach in the paper, including formulation of GAN inversion, generation process, degradation modeling, and supervision strategies.
- Experimental results on various datasets show the performance of ShapeInversion compared to other methods using metrics such as Chamfer distance and F-score.
- ShapeInversion has applications in computer-aided design, virtual reality, and robotics.

## Summary

ShapeInversion is a novel method for 3D shape completion that utilizes Generative Adversarial Network (GAN) inversion. Unlike previous fully supervised methods, ShapeInversion does not require paired training data and instead leverages a pre-trained GAN to search for a latent code that reconstructs the given partial input into a complete shape. This allows it to capture the rich prior of the generative model and achieve impressive results on the ShapeNet benchmark. The proposed approach outperforms state-of-the-art unsupervised methods and is comparable to supervised methods trained with paired data. It also demonstrates remarkable generalization ability by producing robust results for real-world scans and partial inputs of various forms and incompleteness levels. One key advantage of ShapeInversion is its ability to produce multiple valid complete shapes for ambiguous partial inputs, as well as enabling shape manipulation and interpolation. The authors provide additional details about their approach in the paper including formulation of GAN inversion, generation process, degradation modeling, and supervision strategies. They also present experimental results on various datasets and compare ShapeInversion with other methods in terms of performance metrics such as Chamfer distance and F-score. Overall, ShapeInversion offers an effective solution to 3D shape completion by leveraging pre-trained GANs without requiring paired training data. Its ability to generalize well across different domains and produce multiple plausible completions makes it a valuable tool for applications such as computer-aided design, virtual reality, and robotics.