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 [1].

Decoupling Scattering: Pseudo-Label Guided NeRF for Scenes with Scattering Media

Authors: Mingyang Zhang, Junkang Zhang, Faming Fang, Guixu Zhang

AAAI 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our method can effectively simulate hazy and underwater scenes, accurately decouple the scattering medium from objects, estimate atmospheric parameters, and outperform existing methods in novel view synthesis and image restoration tasks. Experimental results substantiate the efficacy of our method, demonstrating its capability to successfully simulate scenes with scattering media, accurately decouple the scattering medium from objects, estimate atmospheric parameters, and achieve superior performance in image restoration and novel view synthesis tasks.
Researcher Affiliation Academia Mingyang Zhang1, Junkang Zhang1, Faming Fang 1,2 *, Guixu Zhang 1 1School of Computer Science and Technology, East China Normal University 2Shanghai Key Laboratory of Multidimensional Information Processing EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Cyclical Optimization Phases for Ne RF Model
Open Source Code Yes Code https://github.com/MMMazart/Deco Ne RF
Open Datasets Yes We synthesized hazy scene data using eight scenes from the well-known LLFF dataset.
Dataset Splits Yes Additionally, 10% of the images in each scene were set aside as a test set to evaluate the performance of novel view synthesis. ... including 20 images from the Red Sea and 18 from the Caribbean Sea, with three images in each set reserved for novel view synthesis evaluation.
Hardware Specification Yes train each scene for 250k iterations on a single NVIDIA 3090 GPU
Software Dependencies No Our implementation is based on the code released in Mip Ne RF-360 (Barron et al. 2022) and Sea Thru-Ne RF (Levy et al. 2023), they were implemented in Jax (Bradbury et al. 2018), while we re-implemented in Py Torch (Paszke et al. 2019).
Experiment Setup Yes We maintain the same learning rate and optimization parameters as used in Mip-Ne RF-360 (Barron et al. 2022) and train each scene for 250k iterations on a single NVIDIA 3090 GPU, and we employ the Pseudo-Label Guided and Cyclical Progressive Dimensional Optimization Strategy during the first 50k iterations.