Rad-NeRF: Ray-decoupled Training of Neural Radiance Field

Authors: Lidong Guo, Xuefei Ning, Yonggan Fu, Tianchen Zhao, Zhuoliang Kang, Jincheng Yu, Yingyan (Celine) Lin, Yu Wang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on five datasets demonstrate that Rad-Ne RF can enhance the rendering performance across a wide range of scene types compared with existing single-Ne RF and multi-Ne RF methods.
Researcher Affiliation Collaboration Lidong Guo1 Xuefei Ning1 Yonggan Fu2 Tianchen Zhao1 Zhuoliang Kang3 Jincheng Yu1 Yingyan (Celine) Lin2 Yu Wang1 1Tsinghua University 2Georgia Institute of Technology 3Meituan
Pseudocode No The paper describes methods and equations but does not include any block labeled pseudocode or algorithm.
Open Source Code Yes Code is available at https://github.com/thu-nics/Rad-Ne RF.
Open Datasets Yes We use five datasets from different types of scenes to evaluate our Rad-Ne RF. (1) Object dataset: we take Masked Tanks-And-Temples dataset (Mask TAT) [13] for evaluation... (2) 360-degree inward/outward-facing datasets: we take Tanks-And-Temples (TAT) dataset with unmasked background [13] and Ne RF-360-v2 dataset [2] to evaluate... (3) free shooting-trajectory datasets: we conduct experiments on Free-Dataset [30] and Scan Net dataset [6]...
Dataset Splits No The paper uses various datasets for evaluation but does not explicitly provide the training, validation, and test splits (e.g., percentages or sample counts) for reproducibility.
Hardware Specification Yes We train the Ne RFs for 20k iterations on a single RTX-3090 GPU.
Software Dependencies No Our Rad-Ne RF is built upon Instant-NGP [18] using a third-party Py Torch implementation 3 and costs no more than one hour of training.
Experiment Setup Yes For Instance-NGP and our Rad-Ne RF, we train the Ne RFs for 20k iterations on a single RTX-3090 GPU. We use Adam optimizer with a batch size of 8192 rays and a learning rate decaying from 1 10 2 to 3 10 4. For the weights of the regularization terms in Equation 6, λ1 is set to 1 10 4 on Ne RF-360-v2 and Free dataset, and is set to 5 10 3 on other datasets. We set λ2 to 1 10 2 on all the datasets.