The idea was to study the concept of NeRF. Upon testing it, we found that the training time was excessively long, so we found papers to reduce it, namely TensoRF: Tensorial Radiance Fields

Upon seeing it, we sought ways to build upon the model and thought of testing a new tensor decomposition to the algorithm. We used Tucker Decomposition.

The big changes to the code were made to the file models/TensoRF.py, by creating a new class TensorTucker, with relevant attributes and methods.

The only dataset that was tested was the lego one. The config should already be set to use TensorTucker. But for other scene, the user should change the config by choosing TensorTucker for the model, and the parameters to [low_number] for both density and appearance.

This repository contains a pytorch implementation for the paper: TensoRF: Tensorial Radiance Fields. Our work present a novel approach to model and reconstruct radiance fields, which achieves super fast training process, compact memory footprint and state-of-the-art rendering quality.

Install environment:

conda create -n TensoRF python=3.8
conda activate TensoRF
pip install torch torchvision
pip install tqdm scikit-image opencv-python configargparse lpips imageio-ffmpeg kornia lpips tensorboard

• Synthetic-NeRF
• Synthetic-NSVF
• Tanks&Temples
• Forward-facing

The training script is in train.py, to train a TensoRF:

python train.py --config configs/lego.txt

we provide a few examples in the configuration folder, please note:

dataset_name, choices = ['blender', 'llff', 'nsvf', 'tankstemple'];

shadingMode, choices = ['MLP_Fea', 'SH'];

model_name, choices = ['TensorVMSplit', 'TensorCP'], corresponding to the VM and CP decomposition. You need to uncomment the last a few rows of the configuration file if you want to training with the TensorCP model；

n_lamb_sigma and n_lamb_sh are string type refer to the basis number of density and appearance along XYZ dimension;

N_voxel_init and N_voxel_final control the resolution of matrix and vector;

N_vis and vis_every control the visualization during training;

You need to set --render_test 1/--render_path 1 if you want to render testing views or path after training.

More options refer to the opt.py.

https://1drv.ms/u/s!Ard0t_p4QWIMgQ2qSEAs7MUk8hVw?e=dc6hBm

python train.py --config configs/lego.txt --ckpt path/to/your/checkpoint --render_only 1 --render_test 1 

You can just simply pass --render_only 1 and --ckpt path/to/your/checkpoint to render images from a pre-trained checkpoint. You may also need to specify what you want to render, like --render_test 1, --render_train 1 or --render_path 1. The rendering results are located in your checkpoint folder.

You can also export the mesh by passing --export_mesh 1:

python train.py --config configs/lego.txt --ckpt path/to/your/checkpoint --export_mesh 1

Note: Please re-train the model and don't use the pretrained checkpoints provided by us for mesh extraction, because some render parameters has changed.

We provide two options for training on your own image set:

1. Following the instructions in the NSVF repo, then set the dataset_name to 'tankstemple'.
2. Calibrating images with the script from NGP: python dataLoader/colmap2nerf.py --colmap_matcher exhaustive --run_colmap, then adjust the datadir in configs/your_own_data.txt. Please check the scene_bbox and near_far if you get abnormal results.

If you find our code or paper helps, please consider citing:

@INPROCEEDINGS{Chen2022ECCV,
  author = {Anpei Chen and Zexiang Xu and Andreas Geiger and Jingyi Yu and Hao Su},
  title = {TensoRF: Tensorial Radiance Fields},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year = {2022}
}