CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes
Authors: Yang Liu, Chuanchen Luo, Zhongkai Mao, Junran Peng, Zhaoxiang Zhang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our method strikes a promising balance between visual quality, geometric accuracy, as well as storage and training costs. Experimental results confirm our state-of-the-art performance in both geometric quality and efficiency. |
| Researcher Affiliation | Academia | 1 NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences 2 University of Chinese Academy of Sciences 3 Shandong University 4 University of Science and Technology Beijing |
| Pseudocode | No | The paper describes methods using textual descriptions and figures (e.g., Figure 2, Figure 4) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | More live demos and official code implementation are available at our project page: https://dekuliutesla.github.io/City Gaussian V2/. |
| Open Datasets | Yes | We utilize the realistic dataset Gau U-Scene (Xiong et al., 2024) and the synthetic dataset Matrix City (Li et al., 2023a). |
| Dataset Splits | Yes | Each scene comprises over 4,000 training images and more than 450 test images |
| Hardware Specification | Yes | All experiments included in this paper are conducted on 8 A100 GPUs. |
| Software Dependencies | No | The paper mentions software like COLMAP, Depth-Anything-V2, and PyTorch (implicitly) but does not provide specific version numbers for any of these components. |
| Experiment Setup | Yes | We set the gradient scaling factor ω to 0.9 and the pruning ratio to 0.025. For depth distortion loss, we empirically find it harmful to performance, and thus set its weight to default value 0. The weight for LDepth is exponentially decayed from 0.5 to 0.0025 during both the pretraining and finetuning stages. LNormal is activated after 7,000 iterations in pretraining and from the beginning in the parallel tuning. Besides, we found that the original normal supervision was overly aggressive for complex scene reconstruction. Consequently, the weight for LNormal is reduced to 0.0125, one-fourth of its original value. We adhere to the default settings in City Gaussian (Liu et al., 2024) for the learning rate and densification schedule. For depth rendering, we utilize median depth for improved geometry accuracy, and for mesh extraction, we employ 2DGS s TSDF-based algorithm with a voxel size of 1m and SDF truncation of 4m. Additionally, Gau U-Scene applies depth truncation of 250m, while Matrix City uses 500m. For vectree quantization, we set the codebook size to 8192 and the quantization ratio to 0.4. |