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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Plenodium: Underwater 3D Scene Reconstruction with Plenoptic Medium Representation
Authors: Changguang WU, Jiangxin Dong, Chengjian Li, Jinhui Tang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on real-world underwater datasets demonstrate that our method achieves significant improvements in 3D reconstruction. Furthermore, we construct a simulated dataset with GT and the controllable scattering medium to demonstrate the restoration capability of our method in underwater scenarios. |
| Researcher Affiliation | Academia | 1Nanjing University of Science and Technology, 2Nanjing Forestry University EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Pseudo-Depth Gaussian Complementation Input: The set of the input cameras V, and corresponded images C; The COLMAP initialized Gaussian primitives, G; Output: The final Gaussian primitives, G ; |
| Open Source Code | Yes | https://plenodium.github.io/ and Justification: Experiment (Sec. 5), Supplementary Material (Sec. B), and our project website: https://plenodium.github.io/ |
| Open Datasets | Yes | Our simulated dataset. We utilize Blender to simulate a dataset with precise GT for restoration evaluation. The dataset includes two scenes (beach and street), each degraded by two types of media (fog [49] and water) at three incremental intensity levels (easy, medium, and hard), yielding 12 systematically structured subsets. Each subset contains 100 images at a resolution of 512 512 pixels. We split them evenly into 50 training and 50 testing samples to ensure a balanced evaluation. and Seathru-Ne RF dataset. The Sea Thru-Nerf dataset [2] includes four real underwater scenes: IUI3 Red Sea, Curaçao, Japanese Gardens Red Sea, and Panama. |
| Dataset Splits | Yes | Each subset contains 100 images at a resolution of 512 512 pixels. We split them evenly into 50 training and 50 testing samples to ensure a balanced evaluation. and There are 29, 20, 20, and 18 images in each scene, respectively, where 25, 17, 17, and 15 images are used for training and the rest of the images are for testing. |
| Hardware Specification | Yes | All the experiments are conducted on an RTX 4090 GPU. |
| Software Dependencies | Yes | We employ the Depth Anything Model [12, 13] as an external image depth estimator to generate the pseudo-depth maps. We use the latest version, V2, and the largest model variant, Vi T-L, which is pretrained on diverse datasets and applied in inference mode without further fine-tuning. |
| Experiment Setup | Yes | During training, we empirically set λL1 = 0.8, λssim = 0.2, and λdepth = 5 in Eqn. 12. The patch number N is set to 16. The maximum degree of the SH coefficients for the medium and Gaussian primitives is set as 3. and We train our model using a volumetric extension of 3D Gaussian Splatting. For reconstruction tasks, we train for 15,000 steps, while for restoration tasks, which require higher accuracy, we extend training to 30,000 steps. Following the progressive training strategy introduced in 3DGS [9], training begins at 1/4 resolution and gradually doubles every 3,000 steps to increase spatial detail. To prevent unstable updates in the early training phase, we apply a 500-step warm-up before the Gaussian refinement. After warm-up, Gaussian refinement (including densification and culling) is performed every 100 steps. and Each parameter group is optimized using the Adam optimizer with ϵ = 10 15 and exponential decay scheduling. For instance, the 3D means are trained with an initial learning rate of 1.6 10 4, which decays to 5 10 5 over time, while opacities are optimized using a fixed learning rate of 0.05. Additional learning rates and scheduler configurations details are provided in Tab. 7. |