An unofficial pytorch implementation of MeRF: Memory-Efficient Radiance Fields for Real-time View Synthesis in Unbounded Scenes.

We support exporting almost lossless baked assets for real-time webGL rendering.

git clone https://github.com/ashawkey/torch-merf.git
cd torch-merf

pip install -r requirements.txt

By default, we use load to build the extension at runtime. However, this may be inconvenient sometimes. Therefore, we also provide the setup.py to build each extension:

# install all extension modules
bash scripts/install_ext.sh

# if you want to install manually, here is an example:
cd raymarching
python setup.py build_ext --inplace # build ext only, do not install (only can be used in the parent directory)
pip install . # install to python path (you still need the raymarching/ folder, since this only install the built extension.)

• Ubuntu 22 with torch 1.12 & CUDA 11.6 on a V100.

We majorly support COLMAP dataset like Mip-NeRF 360. Please download and put them under ./data.

For custom datasets:

# prepare your video or images under /data/custom, and run colmap (assumed installed):
python scripts/colmap2nerf.py --video ./data/custom/video.mp4 --run_colmap # if use video
python scripts/colmap2nerf.py --images ./data/custom/images/ --run_colmap # if use images

First time running will take some time to compile the CUDA extensions.

## train
# mip-nerf 360
python main.py data/bonsai/ --workspace trial_bonsai --enable_cam_center --downscale 4
# front-facing
python main.py data/nerf_llff_data/fern --workspace trial_fern --downscale 4
# nerf-like (need to specify --scale manually)
python main.py data/fox/ --workspace trial_fox --data_format nerf --scale 0.3


## test (export video and baked assets)
python main.py data/bonsai/ --workspace trial_bonsai --enable_cam_center --downscale 4 --test
# the default baking can be very slow (30 min+): it renders all images at full resolution from the training dataset. Use --fast_baking to speed up (just ~1min) at the cost of possibily missing some background blocks:
python main.py data/bonsai/ --workspace trial_bonsai --enable_cam_center --downscale 4 --test --test_no_video --fast_baking

## web renderer
# use VS Code to host the folder and open ./renderer/renderer.html
# follow the instructions and add the baked assets path as URL parameters to start rendering.
# for example:
http://localhost:5500/renderer/renderer.html?dir=../trial_bonsai/assets
http://localhost:5500/renderer/renderer.html?dir=../trial_bonsai/assets&quality=low # phone, low, medium, high

## dense depth supervision (experimental)
cd depth_tools
bash download_models.sh
cd ..

python depth_tools/extract_depth.py --in_dir data/room/images_4 --out_dir data/room/depths

python main.py data/room/ --workspace trial_room --enable_cam_center --downscale 4 --enable_dense_depth

Please check the scripts directory for more examples on common datasets, and check main.py for all options.

The web renderer is slightly modified from the official version, so it is not compatible with the original assets.

• Frequency encoding convention (viewdependency.glsl):

  # original
  x, sin(x), sin(2x), sin(4x), ..., cos(x), cos(2x), cos(4x), ...
  # current
  x, sin(x), cos(x), sin(2x), cos(2x), sin(4x), cos(4x), ...

• Interpolation alignment of sparse grid (fragment.glsl):

  // original
  vec3 posSparseGrid = (z - minPosition) / voxelSize - 0.5;
  // current
  vec3 posSparseGrid = (z - minPosition) / voxelSize;

The baking can be lossless since the baked assets' resolution is the same as the network's resolution, but interpolation must happen after all non-linear functions (i.e., MLP).

This makes the usual hashgrid + MLP combination invalid as

$$ \text{MLP}(\sum_i(w_i * x_i)) \ne \sum_i(w_i * \text{MLP}(x_i)) $$

(using a single linear layer should be able to work though? but the paper uses 2 layers with 64 hidden dims...)

In this implementation we have to manually perform bilinear/trilinear interpolation in torch, and query the 4/8 corners of the grid for each sampling point, which is quite inefficient...

OpenGL's texture() behaves like the F.interpolate(..., align_corners=False). It seems the sparse grid uses align_corners=True convention, while the triplane uses align_corners=False convention... but maybe I'm wrong somewhere, since I have to modify the web renderer to make it work.

The default API is slightly modified for convenience, we need to pass in values in the range of [0, 1] (the bound parameter is removed).

This implmentation (v.s. paper): Indoor 27.38 (v.s. 27.80), Ourdoor 23.41 (v.s. 23.19)

Ours are tested on a V100. Please check the commands under scripts/ to reproduce.

The original paper:

@article{reiser2023merf,
  title={MERF: Memory-Efficient Radiance Fields for Real-time View Synthesis in Unbounded Scenes},
  author={Reiser, Christian and Szeliski, Richard and Verbin, Dor and Srinivasan, Pratul P and Mildenhall, Ben and Geiger, Andreas and Barron, Jonathan T and Hedman, Peter},
  journal={arXiv preprint arXiv:2302.12249},
  year={2023}
}