The Institute of Automation of the Chinese Academy of Sciences proposed an efficient 3D scene reconstruction algorithm citygaussianv2 to overcome the problem of large-scale scene reconstruction!
The AIxiv column continues to report on the world’s top AI research results, and the CityGaussianV2 paper has been selected by ICLR 2025Received, the code is open source. Are you still struggling with time-consuming, memory, and accuracy assessments for large-scale scene reconstruction? CityGaussianV2 provides an efficient and precise solution for fast training, compression, and realistic real-time rendering.
Challenges and Breakthroughs:
Existing algorithms such as 3DGS and its improved versions (SuGaR, 2DGS, GOF) have many problems in large-scale scene reconstruction: the training time is too long, the video memory occupies a huge amount of memory, the accuracy of the reconstruction geometry is insufficient, and there is a lack of effective accuracy evaluation methods. CityGaussianV2 has been thoroughly researched and optimized for these issues.
- Thesis Title: CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes
- Project Homepage: https://www.php.cn/link/b666545e24ea289be13796baae7463e3
- Paper link: https://www.php.cn/link/ca026717248e3b3915f2b15b38f69de6
- Code Links (600+⭐): https://www.php.cn/link/8c2c809f0f90fb60826f6fe06add8fe9
Core improvements in CityGaussian V2:
Based on CityGaussian’s submodel division and data allocation scheme, CityGaussianV2 adopts 2DGS as the basic unit and introduces the following key technologies:
- Densening techniques based on elongation filtration and gradient decoupling: It effectively suppresses the blurring artifacts and excessive primitive growth problems that 2DGS is prone to in complex scenes, and improves the training stability.
- Deep regression supervision: The accuracy of geometric reconstruction is further improved.
- Merge training and compression processes: An end-to-end training process is built, significantly reducing training and compression time.
Experimental results:
CityGaussianV2 achieves significantly better reconstruction results than existing algorithms on multiple datasets, performing well in terms of geometric accuracy and rendering quality, while greatly reducing training time and memory consumption.
Conclusion:
CityGaussianV2 provides an efficient and accurate solution for 3D reconstruction of large-scale complex scenes, and its innovative technology and excellent performance have made important contributions to the development of this field.
That’s all for ICLR 2025Efficient reconstruction of geometrically accurate large-scale complex three-dimensional scenes, the Chinese Academy of Sciences proposed the detailed content of CityGaussianV2, for more information, please pay attention to other related articles on the php Chinese website!
