GaussianBlock: Building Part-Aware Compositional and Editable 3D Scene by Primitives and Gaussians
Authors: Shuyi Jiang, Qihao Zhao, Hossein Rahmani, De Wen Soh, Jun Liu, Na Zhao
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental setup is detailed in Sec.7.1 of the Appendix. 4.1 MAIN RESULTS Primitives Reconstruction. Experiments are conducted on DTU scenes to compare the reconstruction performance of our learnt primitives with SOTA 3D decomposition methods (Liu et al., 2022; Monnier et al., 2023). 4.2 ABLATION STUDY |
| Researcher Affiliation | Academia | Shuyi Jiang1, Qihao Zhao1, Hossein Rahmani2, De Wen Soh1, Jun Liu2, Na Zhao1 1 Singapore Univeristy of Technology and Design, 2 Lancaster University EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Attention-guided Centering Loss Algorithm 2 Split Algorithm 3 Fuse |
| Open Source Code | Yes | Our reconstructed scenes are evidenced to be disentangled, compositional, and compact across diverse benchmarks, enabling seamless, direct and precise editing while maintaining high quality. The code is available at https://github.com/Jiangshuyi0V0/Gaussian Block. |
| Open Datasets | Yes | Thanks to these innovative designs, our method demonstrates state-of-the-art part-level decomposition and controllable, precise editability, with competitive fidelity across various benchmarks, including DTU (Jensen et al., 2014), Nerfstudio (Tancik et al., 2023), Blended MVS (Yao et al., 2020), Mip-360-Garden (Barron et al., 2021) and Tank&Temple-Truck (Knapitsch et al., 2017). |
| Dataset Splits | No | The paper mentions using specific datasets like DTU, Nerfstudio, Blended MVS, Mip-360-Garden, and Tank&Temple-Truck, and conducts experiments on DTU scenes, but it does not explicitly specify exact training/test/validation splits (e.g., percentages, sample counts, or specific predefined split instructions) for these datasets within the provided text. |
| Hardware Specification | Yes | The overall training time is around 6 hours on a single 4090Ti. |
| Software Dependencies | No | The paper mentions using Adam for parameter optimization and implies the use of the 3D Gaussian Splatting framework (Kerbl et al., 2023) but does not provide specific version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | In the primitive optimization step, the initial K is set to 10, while the input image is resized to 400 300 as (Monnier et al., 2023). Approximately 50k iterations are performed for each scene. Besides, the splitting threshold β, the fusion warm up steps ξ and the weight γ of the LAC are set to 0.7, 100 and 0.2, respectively. For the Gaussian Splatting, we use Adam for parameter optimization (the same hyperparameter values are used across all subjects). We set the learning rate to 5e-3 for the position and 1.7e-2 for the scaling of 3D Gaussians and keep the same learning rates as 3D Gaussian Splatting (Kerbl et al., 2023) for the rest of the parameters. We train for 20k iterations, and exponentially decay the learning rate for the splat positions until the final iteration, where it reaches 0.01 the initial value. We enable adaptive density control with binding inheritance every 2,000 iterations. |