Summary of Gtr: Improving Large 3d Reconstruction Models Through Geometry and Texture Refinement, by Peiye Zhuang et al.
GTR: Improving Large 3D Reconstruction Models through Geometry and Texture Refinement
by Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, Sergey Tulyakov, Hsin-Ying Lee
First submitted to arxiv on: 9 Jun 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This novel approach for 3D mesh reconstruction from multi-view images takes inspiration from large reconstruction models like LRM that use a transformer-based triplane generator and a Neural Radiance Field (NeRF) model trained on multi-view images. The method introduces several important modifications to the LRM architecture, enhancing 3D reconstruction quality by improving multi-view image representation and computational efficiency during training. Additionally, meshes are extracted from the NeRF field in a differentiable manner and fine-tuned through mesh rendering for improved geometry reconstruction and supervision at full image resolution. This leads to state-of-the-art performance on both 2D and 3D evaluation metrics, such as a PSNR of 28.67 on the Google Scanned Objects (GSO) dataset. A lightweight per-instance texture refinement procedure is also introduced to fine-tune the triplane representation and NeRF color estimation model on the mesh surface using input multi-view images in just 4 seconds, achieving faithful reconstruction of complex textures like text. The approach enables various downstream applications, including text- or image-to-3D generation. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper proposes a new way to build 3D models from multiple pictures taken from different angles. It’s based on big models that use special generators and training data. But the authors made some changes to make it better. They fixed some problems with the original model, so it works more efficiently and accurately. Then, they took the 3D model and used it to create a new mesh that’s more detailed and realistic. This helped them get even better results on tests like PSNR. The method also has a special trick for making textures look real, like text or pictures. It’s fast too! The authors think this will be useful for making new things, like turning text into 3D models. |
Keywords
» Artificial intelligence » Transformer
