Splannequin: Freezing Monocular Mannequin-Challenge Footage with Dual-Detection Splatting

(Top) A monocular Mannequin-Challenge video is intended to resemble large-scale frozen frames, yet real-world recordings inevitably contain slight body movements. The red crops across successive frames (Ii) highlight these noticeable movements. (Bottom) After our processing, every crop (green boxes) of successive frames remains perfectly static. Splannequin analyzes the entire video and resynthesizes a temporally consistent sequence of views at t* while preserving overall visual fidelity.

Please follow the 3D-GS to install the relative packages.

git clone https://github.com/chien90190/splannequin-gs.git
cd splannequin-gs
mkdir models
git submodule update --init --recursive
conda create -n splannequin python=3.8 -y
conda activate splannequin

module load cuda/11.7
export CC=/usr/bin/gcc-10
export CXX=/usr/bin/g++-10

pip install -r requirements.txt
pip install -e submodules/depth-diff-gaussian-rasterization
pip install -e submodules/simple-knn

Download EvaluationNet.pth.tar following the instructions at ViewEvaluationNet and place it in the ./models/ directory. This model is used for aesthetic evaluation mode.

First, download the dataset metadata from Hugging Face to get dataset.csv.

Then, run the scripts to download frames and generate 3D reconstruction data:

# 1. Download frames from YouTube
python scripts/download.py

# 2. (Optional) Run COLMAP
python scripts/colmap.py ./data/splannequin --gpu

Due to YouTube encoding changes, downloaded frames may have slight temporal jitter compared to the original research data. We strongly recommend running the COLMAP script above to generate aligned camera poses, replacing any provided metadata.

The script will automatically organize the data into the required format:

video_folder/
├── images/       <-- Input
├── sparse/       <-- Output
├── database.db   <-- Output

To process your own videos, organize your data as follows:

my_video/
└── images/
    ├── 0001.png
    └── 0002.png

Then run the reconstruction script pointing to the parent folder:

python scripts/colmap.py path/to/my_video --gpu

For training a scenes such as 1.1 scene, run

bash scripts/train.sh ./data/splannequin/1.1

Run the following script to render the images. We provide several modes for rendering:

• normal: renders the scene as a video with regular time running.
• fixed-time-view: fix the entire scene in a single timestamp.
• time-sweeping-view: select a view and then run the timestamps.
• fixed-time-smooth: fix the entire scene in a single timestamp with a interpolated original trajectory.

# to run all timestamp indicies as the fixed time for the scene (every 8th timestamps)
bash scripts/render.sh ./exp/splannequin/1.1 

# to run selected timestamp indicies (every index corresponds to every 8th frames)
bash scripts/render.sh ./exp/splannequin/1.1 0,1,2,3,4

You can just run the following script to evaluate the model. It runs aesthetic (ViewEvaluationNet), musiq, topiq_nr, hyperiqa, and clipiqa.

bash scripts/eval.sh ./exp/splannequin/1.1

Follow the instruction in COVER to process the videos.

This project is still under development. Please feel free to raise issues or submit pull requests to contribute to our codebase.

Some source code of ours is borrowed from 3DGS, K-planes, HexPlane, TiNeuVox, Depth-Rasterization, 4DGaussians. We sincerely appreciate the excellent works of these authors.

@InProceedings{chien2026splannequin,
  author    = {Hao-Jen Chien, Yi-Chuan Huang, Chung-Ho Wu, Wei-Lun Chao, Yu-Lun Liu},
  title     = {Splannequin: Freezing Monocular Mannequin-Challenge Footage with Dual-Detection Splatting},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  month     = {March},
  year      = {2026},
}