Recovering Dynamic 3D Sketches from Videos

CVPR 2025

Jaeah Lee · Changwoon Choi · Young Min Kim · Jaesik Park

   

TL;DR: Liv3Stroke🖊️ reconstructs dynamic sketches with deformable 3D strokes directly from video frames!

🔧 Installation

# Create the conda envrionment
sh scripts/install.sh
# Find and install an appropriate version of pytorch3d
python scripts/install_pytorch3d.py

🖌️ How to Draw

Train

# Train from start
python main.py --config ours/lego.yaml -ep sanity_check -eg video -en lego -ev 0
# Train from the checkpoint
python main.py --config ours/lego.yaml -ep sanity_check -eg video -en lego -ev 0 -ck [checkpoint] --resume

If you prefer to run it by directly specifying the hyperparameters in the YAML file, you can try the following:

python main.py --config ours/lego.yaml -ep sanity_check -eg bezier -en lego -ev 0 --data.params.root=./data/custom_data_root --method.eval_gap=20 --method.drawer_params.curve_params.num_strokes=64

Prediction

# Under a moving camera trajectory
python main.py --config logs/sanity_check/bezier/lego/config.yaml -ck logs/sanity_check/bezier/lego/best.ckpt --predict
# Under a fixed viewpoint
python main.py --config logs/sanity_check/bezier/lego/config.yaml -ck logs/sanity_check/bezier/lego/best.ckpt --predict --fixed

💾 Dataset

You can download our dataset from this link.

📧 Contact

If you have any questions, please email hayanz@snu.ac.kr.

🌏 Citation

@inproceedings{lee2025recovering,
    author    = {Lee, Jaeah and Choi, Changwoon and Kim, Young Min and Park, Jaesik},
    title     = {Recovering Dynamic 3D Sketches from Videos},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {12423--12432}
}