Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Bootstraping Clustering of Gaussians for View-consistent 3D Scene Understanding
Authors: Wenbo Zhang, Lu Zhang, Ping Hu, Liqian Ma, Yunzhi Zhuge, Huchuan Lu
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on LERF-Mask, 3D-OVS, and Scan Net datasets demonstrate that Free GS performs comparably to state-of-the-art methods while avoiding the complex data preprocessing workload. |
| Researcher Affiliation | Collaboration | Wenbo Zhang1, Lu Zhang1*, Ping Hu2, Liqian Ma3, Yunzhi Zhuge1, Huchuan Lu1 1Dalian University of Technology 2University of Electronic Science and Technology of China 3ZMO AI |
| Pseudocode | No | The paper describes the methodology in detail using mathematical formulas and prose but does not include any clearly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | Code https://github.com/wb014/Free GS |
| Open Datasets | Yes | To evaluate the effectiveness of our method, we conduct experiments on two widely used datasets: LERFMask (Ye et al. 2023) and 3D-OVS (Liu et al. 2023). ... Except for novelview perception, we evaluate the proposed method on 3D object detection and establish a benchmark on Scan Net (Dai et al. 2017), which is a widely used dataset with complex indoor scenes for 3D scene understanding. |
| Dataset Splits | Yes | For each scene, we sample 10% views for reconstruction, with the sampling strategy based on image sharpness to avoid motion blur. |
| Hardware Specification | Yes | We implement the proposed method based on 3DGS (Kerbl et al. 2023) and conduct all experiments on an Nvidia A800 GPU. |
| Software Dependencies | Yes | We implement the proposed method based on 3DGS (Kerbl et al. 2023) and conduct all experiments on an Nvidia A800 GPU. We use the cu ML library (Raschka, Patterson, and Nolet 2020) to support all machine learning algorithms in our work, such as HDBSCAN and PCA, for GPU acceleration. |
| Experiment Setup | Yes | For each scene, the first training phase takes 30k iterations and the second phase for 7k iterations. We set the hyperparameters of λS and λC as 0.1 and 0.05 respectively, and set the feature vector dimension D = 128. We use the Adam optimizer with a learning rate of 0.0025 and 0.0001 for GS and CNN layers respectively. For LF , we set γ = 0.3. For LS, we set K = 5 and T as 0.1% of the total number of Gaussians. We apply LS and LC every 10 iterations for acceleration. For hyperparameters used in HDBSCAN, we set min samples = 20 in experiments on LERF-Mask, and min samples = 60 in experiments on 3D-OVS and Scan Net. |