Rethinking Open-Vocabulary Segmentation of Radiance Fields in 3D Space
Authors: Hyunjee Lee, Youngsik Yun, Jeongmin Bae, Seoha Kim, Youngjung Uh
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experiments, we demonstrate the superiority of our method regarding 1) 3D and rendered 2D segmentation accuracy, 2) training and rendering time, and 3) consistency across viewpoints. In this section, we evaluate our methods on various datasets and compare them to competitors in terms of 3D and rendered 2D segmentation with given text queries. We choose LERF, Open Ne RF, LEGaussians, and Lang Splat as competitors, which are open-sourced. We re-evalute the competitors using the official code3. Appendix provides details of the datasets. |
| Researcher Affiliation | Academia | Hyunjee Lee*, Youngsik Yun*, Jeongmin Bae, Seoha Kim, Youngjung Uh Yonsei University EMAIL |
| Pseudocode | No | The paper describes methods and processes in textual paragraphs and illustrates concepts with figures (e.g., Figure 2, Figure 3, Figure 4, Figure 5). There are no explicitly labeled 'Pseudocode' or 'Algorithm' blocks or sections. |
| Open Source Code | Yes | Code, checkpoints, and annotations are available at the project page. Project page https://hyunji12.github.io/Open3DRF/ |
| Open Datasets | Yes | We evaluate our methods on various datasets and compare them to competitors in terms of 3D and rendered 2D segmentation with given text queries. We choose LERF, Open Ne RF, LEGaussians, and Lang Splat as competitors, which are open-sourced. We re-evalute the competitors using the official code3. Appendix provides details of the datasets. ... Quantitative Results of Rendered 2D Segmentation on LERF and 3D-OVS datasets. ... Table 1: 3D Segmentation Accuracy Comparison on the Replica dataset. |
| Dataset Splits | No | The paper references datasets such as LERF, 3D-OVS, and Replica for evaluation. However, it does not explicitly provide specific training/test/validation split percentages, sample counts, or detailed methodologies for data partitioning. Footnote 3 mentions 'evaluation views' but does not specify how the overall dataset is split for training and testing. |
| Hardware Specification | Yes | Table 3 shows the computational times of the LERF waldo kitchen scene with a single RTX A5000. We report the training time and rendering time for both ours and the competitors. |
| Software Dependencies | No | The paper mentions several software frameworks and models like Ne RFs, 3DGS, CLIP (Radford et al. 2021), SAM (Kirillov et al. 2023), DINO (Caron et al. 2021), Nerfstudio (Tancik et al. 2023), and Su Ga R (Gu edon and Lepetit 2023). However, it does not specify any version numbers for these software components or libraries, which is required for a reproducible description. |
| Experiment Setup | No | The paper describes the proposed methods, loss functions (e.g., Eq. (1), Eq. (4)), and evaluation protocols. However, it does not provide specific experimental setup details such as concrete hyperparameter values (e.g., learning rate, batch size, number of epochs, optimizer settings), model initialization details, or other system-level training configurations in the main text. |