GaussianCut: Interactive segmentation via graph cut for 3D Gaussian Splatting
Authors: Umangi Jain, Ashkan Mirzaei, Igor Gilitschenski
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical evaluations show the adaptability of Gaussian Cut across a diverse set of scenes. Gaussian Cut achieves competitive performance with state-of-the-art approaches for 3D segmentation without requiring any changes to the underlying Gaussian-based scene representation. Project page: https: //umangi-jain.github.io/gaussiancut/. 38th Conference on Neural Information Processing Systems (Neur IPS 2024). 4 Experimental Setup 5.1 Quantitative results 5.2 Qualitative results 5.3 Ablation and sensitivity study |
| Researcher Affiliation | Academia | Umangi Jain, Ashkan Mirzaei, and Igor Gilitschenski University of Toronto {umangi, ashkan, gilitschenski}@cs.toronto.edu |
| Pseudocode | No | The paper describes the overall pipeline and mathematical formulations but does not include a formal pseudocode block or algorithm section. |
| Open Source Code | Yes | Project page: https: //umangi-jain.github.io/gaussiancut/. Code is available at: https://github.com/umangi-jain/gaussiancut |
| Open Datasets | Yes | Datasets: For quantitative evaluation, we test the scenes from LLFF [34], Shiny [54], SPIn-Ne RF [37], and 3D-OVS [29]. All selected scenes from the LLFF and Shiny datasets are real-world front-facing scenes, with 20-62 images each. SPIn-Ne RF provides a collection of scenes from some of the widely-used Ne RF datasets [16, 26, 34, 35, 56]. It contains a combination of front-facing and 360 inward-facing real-world scenes. 3D-OVS contains scenes featuring long-tail objects. |
| Dataset Splits | No | The paper describes evaluation on test data, but does not specify explicit train/validation/test splits in terms of percentages, sample counts, or citations to predefined splits for its own method or the 3DGS optimization process. It mentions 'training views' for 3DGS optimization but not specific splits. |
| Hardware Specification | Yes | All reported times have been obtained using an NVIDIA RTX 4090 GPU and an Intel Core i9-13900KF CPU. |
| Software Dependencies | No | The paper mentions software like 'SAM-Track [9]' and 'Grounding-DINO [31]' and that it follows 'the code provided by Kerbl et al. [23]', but it does not specify version numbers for any of these software dependencies. |
| Experiment Setup | Yes | For all the evaluations, we resize the longer image size to 1008, as commonly practiced in novelview synthesis. We optimize 3DGS model for each of the considered scenes, without making any changes to the original 3DGS code. For coarse splatting, we keep the cut-off threshold τ = 0.9 for front-facing views and τ = 0.3 for the 360 inward-facing scenes. ... For graph cut, we keep γ = 0.1 for neighboring pairwise position weights and γ = 1 for all other weights, the number of neighbors for every node as 10, and the number of clusters for high-confidence nodes as 4 for sink and 1 for source. λ, λn, λu can be adjusted depending on the scene and the quality of coarse splatting but generally, λn = λu = 1 and λ = 0.5 give decent results. |