Dynamic Fluid Surface Reconstruction Using Deep Neural Network
Project Page | Paper
Simron Thapa, Nianyi Li, Jinwei Ye, Imaging and Vision Lab, Louisiana State University. In CVPR 2020 (oral).
We present a dynamic fluid surface reconstruction network that recovers time-varying 3D fluid surfaces from a single viewpoint.
Contributions
- We design physics-motivated loss functions for network training
- We synthesize a large fluid dataset using physics-based modeling and rendering
- Our network is validated on real captured fluid data
Datasets
Please remember to cite the paper if you use this dataset.
Training and Testing
The data preprocessing code (before training with FSRN-CNN) and data post-processing code for the predictions (before training with FSRN-RNN) will be made available soon.
python FSRN-CNN-train.py
python FSRN-RNN-train.py
Evaluation
The code for evaluating the predictions with ground truth values. We use accuracy and error matrics.
python evaluate_metrics.py
Results
- Synthetic Results
- Real Results
- Re-rendered Results
Citation
@InProceedings{Thapa_2020_CVPR,
author = {Thapa, Simron and Li, Nianyi and Ye, Jinwei},
title = {Dynamic Fluid Surface Reconstruction Using Deep Neural Network},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}