DeblurGS: Gaussian Splatting for Camera Motion Blur

Jeongtaek Oh, Jaeyoung Chung, Dongwoo Lee and Kyoung Mu Lee

IMPORTANT

🚧 Under Construction: Please wait for the cleanup.🚧

TO-DO

Updated Code for reproduction.
Release of ExBlur-NP used in paper (ExBlur Dataset + COLMAP with Blurry Observation Only).
Detailed markdown explanation.

Setup

Local Setup

git clone https://github.com/taekkii/deblurgs.git --recursive
cd deblurgs
conda env create --file environment.yml
conda activate deblurgs

Quickstart

LLFFHOLD Convention Dataset

CUDA_VISIBLE_DEVICES=$gpu python train.py --source_path $PATH_TO_DATASET --model_path $PATH_TO_OUTPUT --eval --resolution $RESOLUTION

Custom Dataset

CUDA_VISIBLE_DEVICES=$gpu python train.py --source_path $PATH_TO_DATASET --model_path $PATH_TO_OUTPUT

Arguments

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).

--resolution / -r

Resolution

--num_subframes

Number of subframes per blurry observation.

--random_init

(Optional) Spread random point clouds within camera frustrums, uniformly.

Camera Motion

--curve_controlpoints_lr

Learning rate for bezier curve control points.

--curve_rotation lr

Learning rate for bezier curve rotation.

--curve_alignment_lr.

Learning rate for alignment parameter

--curve_lr_half_iter.

Learning rate for CURVE_CONTROLPOINTS_LR and CURVE_ROTATION_LR exponentially decays to be half for each CURVE_LR_HALF_ITER.

--curve_start_iter.

Runs original 3DGS pipeline before CURVE_START_ITER.

--curve_order

Bezier curve order.

Important Parameters

--densify_grad_threshold_init / --densify_grad_threshold_init

Gaussian Densification Annealing Strategy: the threshold exponentially decays from DENSIFY_GRAD_THRESHOLD_INIT to DENSIFY_GRAD_THRESHOLD_FINAL

--lambda_t_smooth_init / --lambda_t_smooth_final

Temporal smoothness: regularizes abrupt change between adjacent sub-frame renderings. Exponentially decays from LAMBDA_T_SMOOTH_INIT to LAMBDA_T_SMOOTH_FINAL

--lambda_depth_tv

(Optional) 2D TV loss for depth map. Helpful for smooth geometry, but trades reconstruction quality off.

Field-captured Video.

python ./scripts/run_colmap.py --video_path VIDEO/PATH/video.mp4 --result_path COLMAP_WORKSPACE --focal_length OPTIONAL_FOCAL_LENGTH_IF_KNOWN

Do not include --eval

CUDA_VISIBLE_DEVICES=$gpu python train.py --source_path COLMAP_WORKSPACE --model_path $PATH_TO_OUTPUT