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..
Looking Into the Water by Unsupervised Learning of the Surface Shape
Authors: Ori Lifschitz, Tali Treibitz, Dan Rosenbaum
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To test our method we use three datasets. We use the James Real1 dataset [9], which contains 7 sequences of images acquired in a water tank using 50 fps acquisition rate (examples in Fig. 4). Additionally, we use the Tian Set, which is also a real captured dataset by Tian and Narasimhan [23] using a 125 fps camera. We also generate a synthetic dataset using the method in [20], with 3 wave types, resulting in 11 sequences of images. We compare our method to NDIR [11], which is our unsupervised baseline and to Li et al. [12] which is the state-of-the-art supervised method on single images. In all experiments we use a sequence size of 10 frames (except for the batch-size ablation). We conduct ablation studies to validate benefits of modeling surface-height and spatio-temporal information, to examine design choices and to evaluate the impact of the input sequence size (Sec. 4.3). Additional results and videos are provided in the supplementary material. |
| Researcher Affiliation | Academia | Ori Lifschitz Hatter Department of Marine Technologies Charney School of Marine Sciences, University of Haifa Haifa, Israel https://github.com/Ori Lifschitz/RDR-Su Grad Tali Treibitz Hatter Department of Marine Technologies Charney School of Marine Sciences, University of Haifa Haifa, Israel EMAIL Dan Rosenbaum Department of Computer Science University of Haifa Haifa, Israel EMAIL |
| Pseudocode | No | The paper describes methods in prose and with architectural diagrams (Fig. 2) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | All our code and data will be made available upon publication. [...] We provide code, datasets and scripts to run exeperiments in the supplementary material .zip file. [...] https://github.com/Ori Lifschitz/RDR-Su Grad |
| Open Datasets | Yes | To test our method we use three datasets. We use the James Real1 dataset [9], which contains 7 sequences of images acquired in a water tank using 50 fps acquisition rate (examples in Fig. 4). Additionally, we use the Tian Set, which is also a real captured dataset by Tian and Narasimhan [23] using a 125 fps camera. We also generate a synthetic dataset using the method in [20], with 3 wave types, resulting in 11 sequences of images. [...] We provide code, datasets and scripts to run exeperiments in the supplementary material .zip file. |
| Dataset Splits | No | The paper mentions using sequences of images for experiments and conducting ablations averaged over the entire dataset, but it does not specify explicit training, validation, or test dataset splits (e.g., percentages or sample counts) for model training or evaluation. |
| Hardware Specification | No | More details on the configurations of the hyperparameters, the amount of iterations on each training stage, hardware setup, memory usage, runtime, and reproducibility scripts are provided in the supplementary material ("Implementation Details"). |
| Software Dependencies | No | Our implementation is based on the code of [11]. In all experiments we use a 2-layer network for Hθ and a 3-layer network for Iϕ, both trained with the Adam optimizer. |
| Experiment Setup | Yes | In all experiments we use a 2-layer network for Hθ and a 3-layer network for Iϕ, both trained with the Adam optimizer. The input to Iϕ is augmented with random Fourier positional encoding with the bandwidth factor set to 8. We use two sets of hyperparameters. One set for both real datasets (Real1 and Tian Set) and the other for the synthetic dataset. More details on the configurations of the hyperparameters, the amount of iterations on each training stage, hardware setup, memory usage, runtime, and reproducibility scripts are provided in the supplementary material ("Implementation Details"). |