Sigma-TensoRF

Installing Requirements

''' conda create -n SigmaRF python==3.8
conda activate SigmaRF
pip install -r requirements.txt '''

Datasets

Please arrange the datasets in a folder named "data". The following are the download links to the datasets:

• NeRF Synthetic (Link to Google Drive)
• LLFF (Link to Google Drive)
• Tanks & Temples (Download will start after clicking on link)

We note that we use a processed version of Tanks & Temples, and the intrinsics.txt file of "Ignatius" was incompatible with our code. Please manually replace it with the following contents:

1166.564936839068 0.0 960.0 0.0
0.0 1166.564936839068 540.0 0.0
0.0 0.0 1.0 0.0
0.0 0.0 0.0 1.0

After installation, the structure of the data directory should be as follows:

Click to Unfold

data/  
├─ nerf_llff_data/  
│  ├─ fern/  
│  ├─ flower/  
│  ├─ fortress/  
│  ├─ horns/  
│  ├─ leaves/  
│  ├─ orchids/  
│  ├─ room/  
│  ├─ trex/  
├─ nerf_synthetic/  
│  ├─ chair/  
│  ├─ drums/  
│  ├─ ficus/  
│  ├─ hotdog/  
│  ├─ lego/  
│  ├─ materials/  
│  ├─ mic/  
│  ├─ ship/  
├─ TanksAndTemple/  
│  ├─ Barn/  
│  ├─ Caterpillar/  
│  ├─ Family/  
│  ├─ Ignatius/  
│  ├─ Truck/  

Training

Models are trained with config files using the following command:

python train.py --config configs/nerf_synthetic/hotdog.txt

The variable datadir in the config file controls which scene is used. expname controls the folder in which the results are stored.

The hyper-parameters upsilon ($\upsilon$) and eta ($\eta$) can be controlled from the file train.py.

To control which after-slimming ranks to test, use the test_slimmed variable in train.py.

Results

Please create an empty folder log to store the results. Results of experiments are be stored in log/[expname].

The model (in its full form) is stored in log/[expname].th. For an actually applicable model, do the following steps:

1. Extract the state_dict from the .th file.
2. Remove all masks (density_line_mask, app_line_mask, density_plane_mask, app_plane_mask).
3. Convert the remaining parameter tensors to torch.float16. (Optional)

Then, when slimming is needed, directly truncate all tensors along their rank dimension.

2D synthesis testing results are directly given in the log folder. For rendering use the following:

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

For mesh extraction use the following:

python train.py --config configs/nerf_synthetic/hotdog.txt --ckpt path/to/checkpoint --export_mesh 1

Acknowledgements

Our code is partially based on the codebase of TensoRF. We would like to thank the authors of that work.