Hongrui Cai · Wanquan Feng · Xuetao Feng · Yan Wang · Juyong Zhang
Paper | Project Page | Poster
We propose Neural-DynamicReconstruction (NDR), a template-free method to recover high-fidelity geometry, motions and appearance of a dynamic scene from a monocular RGB-D camera.
The data is organized as NeuS
<case_name>
|-- cameras_sphere.npz # camera parameters
|-- depth
|-- # target depth for each view
...
|-- image
|-- # target RGB each view
...
|-- mask
|-- # target mask each view (For unmasked setting, set all pixels as 255)
...
Here cameras_sphere.npz follows the data format in IDR, where world_mat_xx denotes the world-to-image projection matrix, and scale_mat_xx denotes the normalization matrix.
You can download a part of pre-processed KillingFusion data here and unzip it into ./.
Important Tips: If the pre-processed data is useful, please cite the related paper(s) and strictly abide by related open-source license(s).
Clone this repository and create the environment (please notice CUDA version)
git clone https://github.com/USTC3DV/NDR-code.git
cd NDR-code
conda env create -f environment.yml
conda activate ndrDependencies (click to expand)
- torch==1.8.0
- opencv_python==4.5.2.52
- trimesh==3.9.8
- numpy==1.21.2
- scipy==1.7.0
- PyMCubes==0.1.2
- Training
python train_eval.py- Evaluating pre-trained model
python pretrained_eval.py