DGD: Dynamic 3D Gaussians Distillation

Isaac Labe, Noam Issachar, Itai Lang, Sagie Benaim
| Webpage | Full Paper | arXiv |

Abstract

We tackle the task of learning dynamic 3D semantic radiance fields given a single monocular video as input. Our learned semantic radiance field captures per-point semantics as well as color and geometric properties for a dynamic 3D scene, enabling the generation of novel views and their corresponding semantics. This enables the segmentation and tracking of a diverse set of 3D semantic entities, specified using a simple and intuitive interface that includes a user click or a text prompt. To this end, we present DGD, a unified 3D representation for both the appearance and semantics of a dynamic 3D scene, building upon the recently proposed dynamic 3D Gaussians representation. Our representation is optimized over time with both color and semantic information. Key to our method is the joint optimization of the appearance and semantic attributes, which jointly affect the geometric properties of the scene. We evaluate our approach in its ability to enable dense semantic 3D object tracking and demonstrate high-quality results that are fast to render, for a diverse set of scenes.

Pipeline

Dataset

In our paper, we use:

• synthetic dataset from D-NeRF.
• real-world dataset from Hyper-NeRF.

We organize the datasets as follows:

├── data
│   | D-NeRF 
│     ├── hook
│     ├── standup 
│     ├── ...
│   | HyperNeRF
│     ├── interp
│     ├── misc
│     ├── vrig

Setup

Environment

git clone https://github.com/Isaaclabe/DGD-Dynamic-3D-Gaussians-Distillation.git --recursive
cd DGD-Dynamic-3D-Gaussians-Distillation

conda create -n DGD_env python=3.7
conda activate DGD_env

# install pytorch
pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116

# install dependencies
pip install -q plyfile
pip install git+https://github.com/openai/CLIP.git
pip install timm
pip install -r requirements.txt

Setup the submodules

To run the training and rendering code, you need to setup the rasterizer and the Lseg-CLIP model. This is done by using the following instruction,

# The following part setup the gaussian rasterizer module:

cd DGD-Dynamic-3D-Gaussians-Distillation/submodules/diff-gaussian-rasterization
python setup.py build_ext
mkdir DGD-Dynamic-3D-Gaussians-Distillation/diff_gaussian_rasterization
mv DGD-Dynamic-3D-Gaussians-Distillation/submodules/diff-gaussian-rasterization/build/lib.linux-x86_64-cpython-310/diff_gaussian_rasterization/_C.cpython-310-x86_64-linux-gnu.so DGD-Dynamic-3D-Gaussians-Distillation/diff_gaussian_rasterization
mv DGD-Dynamic-3D-Gaussians-Distillation/submodules/diff-gaussian-rasterization/diff_gaussian_rasterization/__init__.py DGD-Dynamic-3D-Gaussians-Distillation/diff_gaussian_rasterization

# The following part setup the simple knn module:

cd DGD-Dynamic-3D-Gaussians-Distillation/submodules/simple-knn
python setup_knn.py build_ext
mkdir DGD-Dynamic-3D-Gaussians-Distillation/simple_knn
mv DGD-Dynamic-3D-Gaussians-Distillation/submodules/simple-knn/build/lib.linux-x86_64-cpython-310/simple_knn/_C.cpython-310-x86_64-linux-gnu.so DGD-Dynamic-3D-Gaussians-Distillation/simple_knn

# The following part setup the Lseg module:

cd DGD-Dynamic-3D-Gaussians-Distillation/lseg_minimal
python setup.py build develop
rm -rf DGD-Dynamic-3D-Gaussians-Distillation/lseg_minimal/lseg
mv DGD-Dynamic-3D-Gaussians-Distillation/lseg_minimal/build/lib/lseg/ DGD-Dynamic-3D-Gaussians-Distillation/lseg_minimal
rm -rf DGD-Dynamic-3D-Gaussians-Distillation/lseg_minimal/build

Train

Use the DINOv2 foundation model

To run the optimizer using the DINOv2 foundation model, simply use

python train.py -s path/to/your/dataset -m output/exp-name --fundation_model "DINOv2" --semantic_dimension 384

Command Line Arguments for train.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--Lseg_model_path

The path where the pre-trained Lseg minimal model should be stored.

--fundation_model

The 2D foundation model used for semantic features. Options are "DINOv2" or "Lseg_CLIP" ("DINOv2" by default).

--semantic_dimension

The dimension of the semantic feature, which is 384 for the DINOv2 model and 512 for the Lseg-CLIP model (384 by default).

--loss_reduce

The factor by which the semantic loss is reduced, calculated as loss = loss_color + loss_reduce * loss_semantic (0.5 by default).

--semantic_start

The iteration index at which semantic optimization begins (25_000 by default).

--semantic_stop

The iteration index at which semantic optimization stops (40_000 by default).

--stop_MLP

The iteration index at which MLP optimization stops, which should be equal to the semantic_start index (25_000 by default).

--iterations

The total number of iterations for training (40_000 by default).

--warm_up

The iteration index until which MLP optimization is paused at the beginning of the optimization (3000 by default).

--images / -i

Alternative subdirectory for COLMAP images (images by default).

--eval

Add this flag to use a MipNeRF360-style training/test split for evaluation.

--resolution / -r

Specifies resolution of the loaded images before training. If provided 1, 2, 4 or 8, uses original, 1/2, 1/4 or 1/8 resolution, respectively. For all other values, rescales the width to the given number while maintaining image aspect. If not set and input image width exceeds 1.6K pixels, inputs are automatically rescaled to this target.

--data_device

Specifies where to put the source image data, cuda by default, recommended to use cpu if training on large/high-resolution dataset, will reduce VRAM consumption, but slightly slow down training. Thanks to HrsPythonix.

--white_background / -w

Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.

--sh_degree

Order of spherical harmonics to be used (no larger than 3). 3 by default.

--convert_SHs_python

Flag to make pipeline compute forward and backward of SHs with PyTorch instead of ours.

--convert_cov3D_python

Flag to make pipeline compute forward and backward of the 3D covariance with PyTorch instead of ours.

--debug

Enables debug mode if you experience erros. If the rasterizer fails, a dump file is created that you may forward to us in an issue so we can take a look.

--debug_from

Debugging is slow. You may specify an iteration (starting from 0) after which the above debugging becomes active.

--ip

IP to start GUI server on, 127.0.0.1 by default.

--port

Port to use for GUI server, 6009 by default.

--test_iterations

Space-separated iterations at which the training script computes L1 and PSNR over test set, 7000 30000 by default.

--save_iterations

Space-separated iterations at which the training script saves the Gaussian model, 7000 30000 <iterations> by default.

--checkpoint_iterations

Space-separated iterations at which to store a checkpoint for continuing later, saved in the model directory.

--start_checkpoint

Path to a saved checkpoint to continue training from.

--quiet

Flag to omit any text written to standard out pipe.

--feature_lr

Spherical harmonics features learning rate, 0.0025 by default.

--opacity_lr

Opacity learning rate, 0.05 by default.

--scaling_lr

Scaling learning rate, 0.005 by default.

--rotation_lr

Rotation learning rate, 0.001 by default.

--position_lr_max_steps

Number of steps (from 0) where position learning rate goes from initial to final. 30_000 by default.

--position_lr_init

Initial 3D position learning rate, 0.00016 by default.

--position_lr_final

Final 3D position learning rate, 0.0000016 by default.

--position_lr_delay_mult

Position learning rate multiplier (cf. Plenoxels), 0.01 by default.

--densify_from_iter

Iteration where densification starts, 500 by default.

--densify_until_iter

Iteration where densification stops, 15_000 by default.

--densify_grad_threshold

Limit that decides if points should be densified based on 2D position gradient, 0.0002 by default.

--densification_interval

How frequently to densify, 100 (every 100 iterations) by default.

--opacity_reset_interval

How frequently to reset opacity, 3_000 by default.

--lambda_dssim

Influence of SSIM on total loss from 0 to 1, 0.2 by default.

--percent_dense

Percentage of scene extent (0--1) a point must exceed to be forcibly densified, 0.01 by default.

Use the Lseg-CLIP foundation model

To run the optimizer using the Lseg-CLIP foundation model, first download the pre-trained Lseg minimal network model from this link. Once downloaded, you can proceed with the optimizer

python train.py -s path/to/your/dataset -m output/exp-name --Lseg_model_path path/to/your/Lseg-model --fundation_model "Lseg_CLIP" --semantic_dimension 512 --loss_reduce 10

Command Line Arguments for train.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--Lseg_model_path

The path where the pre-trained Lseg minimal model should be stored.

--fundation_model

The 2D foundation model used for semantic features. Options are "DINOv2" or "Lseg_CLIP" ("DINOv2" by default).

--semantic_dimension

The dimension of the semantic feature, which is 384 for the DINOv2 model and 512 for the Lseg-CLIP model (384 by default).

--loss_reduce

The factor by which the semantic loss is reduced, calculated as loss = loss_color + loss_reduce * loss_semantic (0.5 by default).

--semantic_start

The iteration index at which semantic optimization begins (25_000 by default).

--semantic_stop

The iteration index at which semantic optimization stops (40_000 by default).

--stop_MLP

The iteration index at which MLP optimization stops, which should be equal to the semantic_start index (25_000 by default).

--iterations

The total number of iterations for training (40_000 by default).

--warm_up

The iteration index until which MLP optimization is paused at the beginning of the optimization (3000 by default).

--images / -i

Alternative subdirectory for COLMAP images (images by default).

--eval

Add this flag to use a MipNeRF360-style training/test split for evaluation.

--resolution / -r

Specifies resolution of the loaded images before training. If provided 1, 2, 4 or 8, uses original, 1/2, 1/4 or 1/8 resolution, respectively. For all other values, rescales the width to the given number while maintaining image aspect. If not set and input image width exceeds 1.6K pixels, inputs are automatically rescaled to this target.

--data_device

Specifies where to put the source image data, cuda by default, recommended to use cpu if training on large/high-resolution dataset, will reduce VRAM consumption, but slightly slow down training. Thanks to HrsPythonix.

--white_background / -w

Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.

--sh_degree

Order of spherical harmonics to be used (no larger than 3). 3 by default.

--convert_SHs_python

Flag to make pipeline compute forward and backward of SHs with PyTorch instead of ours.

--convert_cov3D_python

Flag to make pipeline compute forward and backward of the 3D covariance with PyTorch instead of ours.

--debug

Enables debug mode if you experience erros. If the rasterizer fails, a dump file is created that you may forward to us in an issue so we can take a look.

--debug_from

Debugging is slow. You may specify an iteration (starting from 0) after which the above debugging becomes active.

--ip

IP to start GUI server on, 127.0.0.1 by default.

--port

Port to use for GUI server, 6009 by default.

--test_iterations

Space-separated iterations at which the training script computes L1 and PSNR over test set, 7000 30000 by default.

--save_iterations

Space-separated iterations at which the training script saves the Gaussian model, 7000 30000 <iterations> by default.

--checkpoint_iterations

Space-separated iterations at which to store a checkpoint for continuing later, saved in the model directory.

--start_checkpoint

Path to a saved checkpoint to continue training from.

--quiet

Flag to omit any text written to standard out pipe.

--feature_lr

Spherical harmonics features learning rate, 0.0025 by default.

--opacity_lr

Opacity learning rate, 0.05 by default.

--scaling_lr

Scaling learning rate, 0.005 by default.

--rotation_lr

Rotation learning rate, 0.001 by default.

--position_lr_max_steps

Number of steps (from 0) where position learning rate goes from initial to final. 30_000 by default.

--position_lr_init

Initial 3D position learning rate, 0.00016 by default.

--position_lr_final

Final 3D position learning rate, 0.0000016 by default.

--position_lr_delay_mult

Position learning rate multiplier (cf. Plenoxels), 0.01 by default.

--densify_from_iter

Iteration where densification starts, 500 by default.

--densify_until_iter

Iteration where densification stops, 15_000 by default.

--densify_grad_threshold

Limit that decides if points should be densified based on 2D position gradient, 0.0002 by default.

--densification_interval

How frequently to densify, 100 (every 100 iterations) by default.

--opacity_reset_interval

How frequently to reset opacity, 3_000 by default.

--lambda_dssim

Influence of SSIM on total loss from 0 to 1, 0.2 by default.

--percent_dense

Percentage of scene extent (0--1) a point must exceed to be forcibly densified, 0.01 by default.

Render

Use the DINOv2 foundation model

To render the segmentation using the DINOv2 foundation model, simply use

python render.py -s path/to/your/dataset -m output/exp-name --fundation_model "DINOv2" --semantic_dimension 384 --iterations 40_000 --frame k --novel_views i --points "(x1,y1)" "(x2,y2)" --thetas "ϴ1" "ϴ2"

Command Line Arguments for render.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--Lseg_model_path

The path where the pre-trained Lseg minimal model should be stored.

--fundation_model

The 2D foundation model used for semantic features. Options are "DINOv2" or "Lseg_CLIP" ("DINOv2" by default).

--semantic_dimension

The dimension of the semantic feature, which is 384 for the DINOv2 model and 512 for the Lseg-CLIP model (384 by default).

--iterations

The total number of iterations for training (40_000 by default).

--frame

Specifies the number of training frames in the dataset.

--novel_views

Command to choose whether to render novel views or training views: if novel_views = -1, training views are rendered; if novel_views = index_of_novel_view, the novel view with the specified index is rendered.

--points

A list of tuples (x, y) representing the coordinates of pixels in the first training frame (similar to clicking on the image).

--thetas

A list of thresholds (float) corresponding to the list of points, used to control the granularity of the segmentation.

--prompt

Text prompt for the Lseg-CLIP segmentation.

Use the Lseg-CLIP foundation model

To render the segentation using the Lseg-CLIP foundation model, first download the pre-trained Lseg minimal network model from this link. Once downloaded, you can proceed with the renderer

python render.py -s path/to/your/dataset -m output/exp-name --Lseg_model_path path/to/your/Lseg-model --fundation_model "Lseg_CLIP" --semantic_dimension 512  --iterations 40_000 --frame k --novel_views i --prompt "text" --thetas "ϴ"

Command Line Arguments for render.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--Lseg_model_path

The path where the pre-trained Lseg minimal model should be stored.

--fundation_model

The 2D foundation model used for semantic features. Options are "DINOv2" or "Lseg_CLIP" ("DINOv2" by default).

--semantic_dimension

The dimension of the semantic feature, which is 384 for the DINOv2 model and 512 for the Lseg-CLIP model (384 by default).

--iterations

The total number of iterations for training (40_000 by default).

--frame

Specifies the number of training frames in the dataset.

--novel_views

Command to choose whether to render novel views or training views: if novel_views = -1, training views are rendered; if novel_views = index_of_novel_view, the novel view with the specified index is rendered.

--points

A list of tuples (x, y) representing the coordinates of pixels in the first training frame (similar to clicking on the image).

--thetas

A list of thresholds (float) corresponding to the list of points, used to control the granularity of the segmentation.

--prompt

Text prompt for the Lseg-CLIP segmentation.

Run the code easily

To simplify training and rendering with this repository, I created a Colab file: DGD-Dynamic-3D-Gaussians-Distillation.ipynb, which can be easily loaded and launched. I recommend using an A100 GPU to significantly reduce computation time during the runs.

BibTex

@misc{labe2024dgd,
      title={DGD: Dynamic 3D Gaussians Distillation}, 
      author={Isaac Labe and Noam Issachar and Itai Lang and Sagie Benaim},
      year={2024},
      eprint={2405.19321},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgement

Our repo is developed based on 3D Gaussian Splatting, DFFs, lseg-minimal and Deformable 3D Gaussians. Many thanks to the authors for opensoucing the codebase.