# clone repository
git clone https://github.com/yzslab/gaussian-splatting-lightning.git
cd gaussian-splatting-lightning
# create virtual environment
conda create -yn gspl python=3.9 pip
conda activate gspl
PyTorch==2.0.1
For CUDA 11.8
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118
pip install -e .
ffmpeg is required if you want to render video: sudo apt install -y ffmpeg
If you want to use nerfstudio-project/gsplat
pip install git+https://github.com/yzslab/gsplat.git
This command will install my modified version, which is required by LightGaussian and Mip-Splatting. If you do not need them, you can also install vanilla gsplat v0.1.12.
If you need SegAnyGaussian
pip install hdbscan scikit-learn==1.3.2 git+https://github.com/facebookresearch/segment-anything.git
For torch==2.0.1
and cuda 11.8:
pip install fvcore iopath
pip install --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py39_cu118_pyt201/download.html
wget -O sam_vit_h_4b8939.pth https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
python main.py fit \
--data.path DATASET_PATH \
-n EXPERIMENT_NAME
It can detect some dataset type automatically. You can also specify type with option --data.parser
. Possible values are: Colmap
, Blender
, NSVF
, Nerfies
, MatrixCity
, PhotoTourism
, SegAnyColmap
, Feature3DGSColmap
.
[NOTE] By default, only checkpoint files will be produced on training end. If you need ply file in vanilla 3DGS's format (can be loaded by SIBR_viewer or some WebGL/GPU based viewer):
python utils/ckpt2ply.py TRAINING_OUTPUT_PATH
, e.g.:
python utils/ckpt2ply.py outputs/lego
python utils/ckpt2ply.py outputs/lego/checkpoints/epoch=300-step=30000.ckpt
--model.save_ply true
python main.py fit \
--viewer \
...
configs/blender.yaml
when training on blender dataset.
python main.py fit \
--config configs/blender.yaml \
...
# the requirements of mask
# * must be single channel
# * zero(black) represent the masked pixel (won't be used to supervise learning)
# * the filename of the mask file must be image filename + '.png',
# e.g.: the mask of '001.jpg' is '001.jpg.png'
... fit \
--data.parser Colmap \
--data.parser.mask_dir MASK_DIR_PATH \
...
You can use utils/image_downsample.py
to downsample your images, e.g. 4x downsample: python utils/image_downsample.py PATH_TO_DIRECTORY_THAT_STORE_IMAGES --factor 4
# it will load images from `images_4` directory
... fit \
--data.parser Colmap \
--data.parser.down_sample_factor 4 \
...
Rounding mode is specified by --data.parser.down_sample_rounding_mode
. Available values are floor
, round
, round_half_up
, ceil
. Default is round
.
Load large dataset without OOM
... fit \
--data.train_max_num_images_to_cache 1024 \
...
Use in your code
from main import cli_main
cli_main(["fit", "--config", "/path/to/config.yaml", "--data.path", "/path/to/data", "args", ...])
### 2.3. Use <a href="https://github.com/nerfstudio-project/gsplat">nerfstudio-project/gsplat</a>
Make sure that command `which nvcc` can produce output, or gsplat will be disabled automatically.
```bash
python main.py fit \
--config configs/gsplat.yaml \
...
[NOTE] Try New Multiple GPU training strategy, which can be enabled during densification.
[NOTE] Multi-GPU training with DDP strategy can only be enabled after densification. You can start a single GPU training at the beginning, and save a checkpoint after densification finishing. Then resume from this checkpoint and enable multi-GPU training.
You will get improved PSNR and SSIM with more GPUs:
# Single GPU at the beginning
python main.py fit \
--config ... \
--data.path DATASET_PATH \
--model.density.densify_until_iter 15000 \
--max_steps 15000
# Then resume, and enable multi-GPU
python main.py fit \
--config ... \
--trainer configs/ddp.yaml \
--data.path DATASET_PATH \
--max_steps 30000 \
--ckpt_path last # find latest checkpoint automatically, or provide a path to checkpoint file
python main.py fit \
--config configs/deformable_blender.yaml \
--data.path ...
python main.py fit \
--config configs/mip_splatting_gsplat.yaml \
--data.path ...
Prune & finetune only currently
Train & densify & prune
... fit \
--config configs/light_gaussian/train_densify_prune-gsplat.yaml \
--data.path ...
Prune & finetune (make sure to use the same hparams as the input model used)
... fit \
--config configs/light_gaussian/prune_finetune-gsplat.yaml \
--data.path ... \
... \
--ckpt_path YOUR_CHECKPOINT_PATH
... fit \
--config configs/gsplat-absgrad.yaml \
--data.path ...
Install diff-surfel-rasterization
first
pip install git+https://github.com/hbb1/diff-surfel-rasterization.git@3a9357f6a4b80ba319560be7965ed6a88ec951c6
Then start training
... fit \
--config configs/vanilla_2dgs.yaml \
--data.path ...
First, train a 3DGS scene using gsplat
python main.py fit \
--config configs/gsplat.yaml \
--data.path data/Truck \
-n Truck -v gsplat # trained model will save to `outputs/Truck/gsplat`
Then generate SAM masks and their scales
Masks
python utils/get_sam_masks.py data/Truck/images
You can specify the path to SAM checkpoint via argument -c PATH_TO_SAM_CKPT
Scales
python utils/get_sam_mask_scales.py outputs/Truck/gsplat
Both the masks and scales will be saved in data/Truck/semantics
, the structure of data/Truck
will like this:
├── images # The images of your dataset
├── 000001.jpg
├── 000002.jpg
...
├── semantic # Generated by `get_sam_masks.py` and `get_sam_mask_scales.py`
├── masks
├── 000001.jpg.pt
├── 000002.jpg.pt
...
└── scales
├── 000001.jpg.pt
├── 000002.jpg.pt
...
├── sparse # colmap sparse database
...
Train SegAnyGS
python seganygs.py fit \
--config configs/segany_splatting.yaml \
--data.path data/Truck \
--model.initialize_from outputs/Truck/gsplat \
-n Truck -v seganygs # save to `outputs/Truck/seganygs`
The value of --model.initialize_from
is the path to the trained 3DGS model
Start the web viewer to perform segmentation or cluster
python viewer.py outputs/Truck/seganygs
Baseline | Partitioning |
---|---|
The implementation here references Block-NeRF, VastGaussian and CityGaussians.
There is no single script to finish the whole pipeline. Please refer to below contents about how to reconstruct a large scale scene.
Training
[NOTE] You must explicitly specify using a gsplat
-based renderer (e.g., configs/gsplat.yaml
, configs/appearance_embedding_renderer
, configs/mip_splatting_gsplat.yaml
, etc) because some parts of the pipeline make assumptions about this. If you do not, it will use diff-gaussian-rasterization
by default, which will cause a drop in the final quality.
First prepare the dataset
# 1. download the dataset
mkdir -p data/MegaNeRF
pushd data/MegaNeRF
wget -O rubble-pixsfm.tgz https://storage.cmusatyalab.org/mega-nerf-data/rubble-pixsfm.tgz
tar -zxf rubble-pixsfm.tgz
popd
# 2. create a colmap sparse model from provided camera poses
wget -O vocab_tree_flickr100K_words256K.bin https://demuc.de/colmap/vocab_tree_flickr100K_words256K.bin
python utils/meganerf2colmap.py data/MegaNeRF/rubble-pixsfm -v vocab_tree_flickr100K_words256K.bin
# 3. down sample images
python utils/image_downsample.py data/MegaNeRF/rubble-pixsfm/colmap/images --factor 3
The intrinsics and extrinsics provided in the Rubble
dataset seem not optimal and will produce a slightly blurry result. Run a sparse reconstruction from scratch if you want to avoid it.
Generate appearance groups
The Rubble dataset contains images in various lighting conditions. Enabling the appearance model can improve the quality. In order to do so, appearance groups must be generated first.
python utils/generate_image_apperance_groups.py \
data/MegaNeRF/rubble-pixsfm/colmap \
--image \
--name appearance_image_dedicated
Partitioning: simply use notebooks/meganerf_rubble_split.ipynb, and remember to change the value of dataset_path
in this notebook
Training
PARTITION_DATA_PATH="data/MegaNeRF/rubble-pixsfm/colmap/partitions-size_60.0-enlarge_0.1-visibility_0.9_0.25"
PROJECT_NAME="MegaNeRF-rubble"
python utils/train_colmap_partitions_v2.py \
${PARTITION_DATA_PATH} \
-p ${PROJECT_NAME} \
--scalable-config utils/scalable_param_configs/appearance.yaml \
--config configs/appearance_embedding_renderer/sh_view_dependent.yaml \
-- \
--data.parser.appearance_groups appearance_image_dedicated \
--model.gaussian.optimization.spatial_lr_scale 15 \
--data.parser.down_sample_factor 3
Parameter Explanations
utils/train_colmap_partitions_v2.py
output_path
in the partitioning notebook-p
: the project name, all the trained partition models will be stored in outputs/PROJECT_NAME
--scalable-config
: the path of the yaml file specifying the hyperparameters that will be adjusted according to the image number--config
: the config file used for training--
: all the parameters after this will be passed to main.py
as-ismain.py
--data.parser.appearance_groups
: the name of the generated appearance groups--model.gaussian.optimization.spatial_lr_scale
: the LR of the 3D means of Gaussians will be multiplied by this value, which determines how finely the structure can be modeled
By default, this value will be calculated automatically according to the camera poses if it is omitted, but this will lead to suboptimal results for large-scale scenes. Therefore you had better provide it manually.
Please note that this is a scene-specific hyperparameter. You should not use the same value for another scene, or even when you run a colmap sparse reconstruction a second time. Additionally, the value 15
is not the optimal one for the dataset Rubble.
--data.parser.down_sample_factor
: down sample factor of the imagesMerging
python utils/merge_partitions_v2.py \
${PARTITION_DATA_PATH} \
-p ${PROJECT_NAME}
Then you can start the web viewer with the merged checkpoint file.
Optional prune and finetune
Prune
python utils/prune_partitions_v2.py \
${PARTITION_DATA_PATH} \
-p ${PROJECT_NAME}
Finetune
PRUNED_PROJECT_NAME="${PROJECT_NAME}-pruned"
python utils/finetune_pruned_partitions_v2.py \
${PARTITION_DATA_PATH} \
-p ${PRUNED_PROJECT_NAME} \
-t ${PROJECT_NAME} \
--scalable-config utils/scalable_param_configs/appearance.yaml
It will load trained model from outputs/${PROJECT_NAME}
, and the finetuned outputs will be saved to outputs/${PRUNED_PROJECT_NAME}
.
Merge finetuned outputs
python utils/merge_partitions_v2.py \
${PARTITION_DATA_PATH} \
-p ${PRUNED_PROJECT_NAME}
LoD Rendering
With LoD, the renderer will select the finer models for partitions close to the camera, and coarser models for those far away.
(a) Preprocess partitions' checkpoints
This is done by simply add an option --preprocess
when running utils/merge_partitions_v2.py
. You need to run it for every level.
# First for the original models
python utils/merge_partitions_v2.py \
--preprocess \
${PARTITION_DATA_PATH} \
-p ${PROJECT_NAME}
# Then for the pruned models
python utils/merge_partitions_v2.py \
--preprocess \
${PARTITION_DATA_PATH} \
-p ${PRUNED_PROJECT_NAME}
(b) Create a LoD config file in YAML
# `data` is the path to your partition data
data: data/MegaNeRF/rubble-pixsfm/colmap/partitions-size_60.0-enlarge_0.1-visibility_0.9_0.25
# `names` is a list of project names, where order represents their detail levels, ranging from fine to coarse
names:
- MegaNeRF-rubble # without pruning
- MegaNeRF-rubble-pruned # pruned
# - MegaNeRF-rubble-pruned_again # add more lines if you have
(c) Start the viewer
python viewer YOU_LOD_CONFIG_FILE_PATH.yaml
Train/prune/finetune multiple partitions in parallel
The options --n-processes
and --process-id
are designed for the parallel purpose. --n-processes
is the total number of GPUs you want to use, which has the same meaning as world size. --process-id
is the unique process id, just like the global rank, but ranging from 1 to the value of --n-processes
here. Please note that both of the options should be placed before --
.
Assuming you have 2 machines, and both of them are equipped with 2 GPUs:
First run:
CUDA_VISIBLE_DEVICES=0 python utils/train_colmap_partitions_v2.py \
... \
--n-processes 4 \
--process-id 1 \
-- \
...
Then run this in another terminal: CUDA_VISIBLE_DEVICES=1 python ... --n-processes 4 --process-id 2
First: CUDA_VISIBLE_DEVICES=0 python ... --n-processes 4 --process-id 3
Then another terminal:CUDA_VISIBLE_DEVICES=1 python ... --n-processes 4 --process-id 4
Or submit partition training jobs to Slurm:
# Adding params after the 2nd `--` means submitting jobs to Slurm,
# and those params belong to `srun`.
# Do not add `--n-processes` and `--process-id` here.
python utils/train_colmap_partitions_v2.py \
... \
-- \
... \
-- \
--gres=gpu:1 # one gpu per-partition training job
Train/finetune a partition with multiple GPUs
With appearance model, the reconstruction quality can be improved when your images have various appearance, such as different exposure, white balance, contrast and even day and night.
This model assign an extra feature vector $\boldsymbol{\ell}^{(g)}$ to each 3D Gaussian and an appearance embedding vector $\boldsymbol{\ell}^{(a)}$ to each appearance group. Both of them will be used as the input of a lightweight MLP to calculate the color.
$$ \mathbf{C} = f \left ( \boldsymbol{\ell}^{(g)}, \boldsymbol{\ell}^{(a)} \right ) $$
Please refer to internal/renderers/gsplat_appearance_embedding_renderer.py for more details.
Baseline | New Model |
---|---|
First generate appearance groups (Colmap or PhotoTourism dataset only)
python utils/generate_image_apperance_groups.py PATH_TO_DATASET_DIR \
--image \
--name appearance_image_dedicated # the name will be used later
The images in a group will share a common appearance embedding. The command above will assign each image a group, which means that will not share any appearance embedding between images.
Then start training
python main.py fit \
--config configs/appearance_embedding_renderer/view_dependent.yaml \
--data.path PATH_TO_DATASET_DIR \
--data.parser Colmap \
--data.parser.appearance_groups appearance_image_dedicated # value here should be the same as the one provided to `--name` above
If you are using PhotoTourism dataset, please replace --data.parser Colmap
with --data.parser PhotoTourism
.
view_independent.yaml
: turn off view dependent effectssh_view_dependent.yaml
: represent view dependent effects using spherical harmonics*-distributed.yaml
: multiple GPUs*-estimated_depth_reg.yaml
/ *-estimated_depth_reg-hard_depth.yaml
: with depth regularizationRemove the dependence on MLP when rendering
It is recommended to use view_independent-*
or sh_view_dependent-*
configs if you want to do so.
By running python utils/fuse_appearance_embeddings_into_shs_dc.py TRAINED_MODEL_DIR
, you can get a fixed appearance checkpoint without requiring a MLP.
submodules/mcmc_relocation
firstpip install submodules/mcmc_relocation
... fit \
--config configs/gsplat-mcmc.yaml \
--model.density.cap_max MAX_NUM_GAUSSIANS \
...
MAX_NUM_GAUSSIANS
is the maximum number of Gaussians that will be used.
Refer to ubc-vision/3dgs-mcmc, internal/density_controllers/mcmc_density_controller.py and internal/metrics/mcmc_metrics.py for more details.
Based on the Appearance Model (2.12.) above, this model can produce a visibility map for every training view indicating whether a pixel belongs to transient objects or not.
The idea of the visibility map is a bit like Ha-NeRF, but rather than uses positional encoding for pixel coordinates, 2D dense grid encoding is used here in order to accelerate training.
Please refer to Ha-NeRF, internal/renderers/gsplat_appearance_embedding_visibility_map_renderer.py
and internal/metrics/visibility_map_metrics.py
for more details.
[NOTE] Though it shows the capability to distinguish the pixels of transient objects, may not be able to remove some artifats/floaters belong to transients. And may also treat under-reconstructed regions as transients.
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch
Download PhotoTourism dataset from here and split file from the "Additional links" here. The split file should be placed at the same path as the dense
directory of the PhotoTourism dataset, e.g.:
├──brandenburg_gate
├── dense # colmap database
├── images
├── ...
├── sparse
...
├── brandenburg.tsv # split file
[Optional] 2x downsize the images: python utils/image_downsample.py data/brandenburg_gate/dense/images --factor 2
python main.py fit \
--config configs/appearance_embedding_visibility_map_renderer/view_independent-2x_ds.yaml \
--data.path data/brandenburg_gate \
-n brandenburg_gate
If you have not downsized images, remember to add a --data.parser.down_sample_factor 1
to the command above.
python main.py validate \
--save_val \
--val_train \
--config outputs/brandenburg_gate/lightning_logs/version_0/config.yaml # you may need to change this path
Then you can find the rendered masks and images in outputs/brandenburg_gate/val
.
This is a bit like a simplified version of Scaling Up 3DGS.
In the implementation here, Gaussians are stored, projected and their colors are calculated in a distributed manner, and each GPU rasterizes a whole image for a different camera. No Pixel-wise Distribution currently.
This strategy works with densification enabled.
[NOTE]
Training
python main.py fit \
--config configs/distributed.yaml \
...
By default, all processes will hold a (redundant) replica of the dataset in memory, which may cause CPU OOM. You can avoid this by adding the option --data.distributed true
, so that each process loads a different subset of the dataset.
Merge checkpoints
python utils/merge_distributed_ckpts.py outputs/TRAINED_MODEL_DIR
python viewer.py outputs/TRAINED_MODEL_DIR/checkpoints/MERGED_CHECKPOINT_FILE
[NOTE] No utilization-based pruning (4.2.3 of the paper) and appearance modeling (4.2.4 of the paper)
Install requirements
pip install diffusers==0.27.2 transformers==4.40.1 scikit-learn
Extract Stable Diffusion features
python utils/sd_feature_extraction.py YOUR_IMAGE_DIR
Training
python main.py fit \
--config configs/spot_less_splats/gsplat-cluster.yaml \
--data.parser.split_mode "reconstruction" \
--data.path YOUR_DATASET_PATH \
-n EXPERIMENT_NAME
python main.py fit \
--config configs/spot_less_splats/gsplat-mlp.yaml \
--data.parser.split_mode "reconstruction" \
--data.path YOUR_DATASET_PATH \
-n EXPERIMENT_NAME
Change the value of --data.parser.split_mode
to keyword
if you are using the RobustNeRF dataset.
Render SLS predicted masks
python utils/render_sls_masks.py outputs/EXPERIMENT_NAME
This is implemented with reference to Hierarchical 3DGS.
Baseline | DepthReg | DepthReg + AppearanceModel |
---|---|---|
Setup Depth Anything V2
# clone the repo.
git clone https://github.com/DepthAnything/Depth-Anything-V2 utils/Depth-Anything-V2
# NOTE: do not run `pip install -r utils/Depth-Anything-V2/requirements.txt`
# download the pretrained model `Depth-Anything-V2-Large`
mkdir utils/Depth-Anything-V2/checkpoints
wget -O utils/Depth-Anything-V2/checkpoints/depth_anything_v2_vitl.pth "https://huggingface.co/depth-anything/Depth-Anything-V2-Large/resolve/main/depth_anything_v2_vitl.pth?download=true"
Dataset pre-processing
python utils/estimate_dataset_depths.py data/Family
Make sure that both the sparse
and images
folders exist in data/Family
.
With the operation above, the structure of data/Family
should be like this:
├── data/Family
├── estimated_depths # generated by `utils/run_depth_anything_v2.py`
├── 00001.jpg.npy
├── ...
├── images
├── 00001.jpg
├── ...
├── sparse # colmap sparse model
├── ...
├── estimated_depth_scales.json # generated by `utils/get_depth_scales.py`
...
Training
python main.py fit \
--config configs/depth_regularization/estimated_inverse_depth-l1.yaml \
--data.path data/Family \
-n EXPERIMENT_NAME
Other available configs:
In my experiments, simply L1 is slightly better than L2 or the one with SSIM.
Per-image metrics will be saved to TRAINING_OUTPUT/metrics
as a csv
file.
python main.py validate \
--config outputs/lego/config.yaml
python main.py test \
--config outputs/lego/config.yaml
python main.py validate \
--config outputs/lego/config.yaml \
--val_train
python main.py <validate or test> \
--config outputs/lego/config.yaml \
--save_val
Then you can find the images in outputs/lego/<val or test>
.
Transform | Camera Path | Edit |
---|---|---|
python viewer.py TRAINING_OUTPUT_PATH
# e.g.:
# python viewer.py outputs/lego/
# python viewer.py outputs/lego/checkpoints/epoch=300-step=30000.ckpt
# python viewer.py outputs/lego/baseline/point_cloud/iteration_30000/point_cloud.ply # only works with VanillaRenderer
python viewer.py \
outputs/garden \
outputs/lego \
outputs/Synthetic_NSVF/Palace/point_cloud/iteration_30000/point_cloud.ply \
--enable_transform
[NOTE] The commands in this section only design for third-party outputs
python viewer.py \
Deformable-3D-Gaussians/outputs/lego \
--vanilla_deformable \
--reorient disable # change to enable when loading real world scene
python viewer.py \
4DGaussians/outputs/lego \
--vanilla_gs4d
# Install `diff-surfel-rasterization` first
pip install git+https://github.com/hbb1/diff-surfel-rasterization.git@28c928a36ea19407cd9754d068bd9a9535216979
# Then start viewer
python viewer.py \
2d-gaussian-splatting/outputs/Truck \
--vanilla_gs2d
python viewer.py \
SegAnyGAussians/outputs/Truck \
--vanilla_seganygs
autonomousvision/mip-splatting
python viewer.py \
mip-splatting/outputs/bicycle \
--vanilla_mip
Q: The viewer shows my scene in unexpected orientation, how to rotate the camera, like the U
and O
key in the SIBR_viewer?
A: Check the Orientation Control
on the right panel, rotate the camera frustum in the scene to the orientation you want, then click Apply Up Direction
.
Besides: You can also click the 'Reset up direction' button. Then the viewer will use your current orientation as the reference.
Q: The web viewer is slow (or low fps, far from real-time).
A: This is expected because of the overhead of the image transfer over network. You can get around 10fps in 1080P resolution, which is enough for you to view the reconstruction quality.