Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, Kwang Moo Yi
arXiv | CVPR | Project Page
This repository contains the implementation and resources for our research paper "Accelerating Neural Field Training via Soft Mining" which was accepted at CVPR 2024.
The paper presents a novel approach to accelerate the training of Neural Fields (NeRFs) by introducing a soft mining strategy. This strategy dynamically selects a subset of rays for each iteration, focusing on those that contribute most to the learning process based on the loss gradient.
For a detailed understanding of the methodology, results, and more, please refer to the full paper.
Please refer to NerfAcc Repository for installation instructions.
After installing Nerfacc, make sure you have NeRF Synthetic or LLFF dataset downloaded.
To run the baseline with Uniform Sampling:
python examples/train_ngp_nerf_prop.py --sampling_type uniform --data_root /path/to/your/dataset --scene scene_name
To run LMC sampling, simply run:
python examples/train_ngp_nerf_prop.py --sampling_type lmc --data_root /path/to/your/dataset --scene scene_name
This project is built upon the work found in Nerfacc Repository. Special thanks to all the contributors of the original repository for laying the groundwork that has enabled us to advance this initiative.
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.
@article{kheradsoftmining2023,
author = {Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, Kwang Moo Yi},
title = {Accelerating Neural Field Training via Soft Mining},
journal = {Arxiv},
year = {2023},
}