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..
I2-NeRF: Learning Neural Radiance Fields Under Physically-Grounded Media Interactions
Authors: Shuhong Liu, Lin Gu, Ziteng Cui, Xuangeng Chu, Tatsuya Harada
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on real-world datasets demonstrate that our method significantly improves both reconstruction fidelity and physical plausibility compared to existing approaches. |
| Researcher Affiliation | Academia | 1The University of Tokyo, 2Tohoku University, 3RIKEN EMAIL |
| Pseudocode | No | The paper describes methods through mathematical formulations and textual descriptions rather than structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available at https://github.com/Shuhong LL/I2-Ne RF. |
| Open Datasets | Yes | Datasets For underwater environments, we use the Sea Thru-Ne RF dataset [45]. For low-light conditions, we conduct evaluations on the LOM dataset [16]. Both datasets follow the original train-test split. In addition, we captured two real-world underwater scenes in Okinawa, Pacific Ocean, using an OLYMPUS Tough TG-6 underwater camera and recorded the corresponding water depths. |
| Dataset Splits | Yes | For the Sea Thru-Ne RF dataset [45], we adopt the common practice in previous studies by assigning every 8th view as the test view. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA RTX A6000 Ada GPU. |
| Software Dependencies | No | Our model is built upon the Zip Ne RF codebase [8], leveraging its hash-based encoding for efficient training. This provides a codebase reference but lacks specific version numbers for ancillary software components like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | In our implementation based on the Zip Ne RF codebase [8], we set the number of proposal sampling points to 128, the number of Ne RF sampling points to 32, and the number of media upsampling points to 32, treating them equally to object samples. The sampling hierarchy level is set to 2. We employ two separate hash grid encoders for object and media components, which introduces a slight increase in training time but is necessary to preserve their distinct geometry and spatial distributions. The output dimensionality of both hash encoders is set to 256. For the LOM dataset [16], we use a batch size of 4096. For the Sea Thru-Ne RF dataset [45], the batch size is set to 2048, and for our captured underwater scenes, it is set to 1024. Each batch size is scaled proportionally to the total number of pixels in the respective dataset. The maximum number of training steps is set to 25,000. We use the Adam optimizer with an initial learning rate of 10-2 and a final learning rate of 10-4. |