We propose a new neural network with Graph-RNN cells, for point cloud sequence prediction
Please cite this paper if you want to use it in your work,
@article{gomes2021spatiotemporal,
title={Spatio-temporal Graph-RNN for Point Cloud Prediction},
author={Pedro Gomes and Silvia Rossi and Laura Toni},
year={2021},
eprint={2102.07482},
archivePrefix={arXiv},
}
Install TensorFlow. The code has been tested with Python 3.6, TensorFlow 1.12.0, CUDA 9.0 and cuDNN 7.21
Compile the code. You must select the correct CUDA version and Tensorflow installed on your computer. For that edit the Makefiles to the paths of your Cuda and Tensorflow directories.
The Makefiles to compile the code are in modules/tf_ops
To train a model to long-term prediction using the GraphRNN
python train-mmnist-GraphRNN.py
To evaluate the model
python eval-mmnist.py
to train the model without color or with color
python train-bodies-GraphRNN.py
python train-bodies-GraphRNN_color.py
to evaluate
python eval-bodies.py
python eval-bodies_color.py
The models were evaluated with the following datasets:
- Moving MNIST Point Cloud (1 digit) 2. Moving MNIST Point Cloud (2 digits) 3. JPEG Dynamic Human Bodies (4000 points)
To create the Human Bodies dataset follow the instruction in the Dataset folder.
The parts of this codebase is borrowed from Related Repos:
- PointRNN TensorFlow implementation: https://github.com/hehefan/PointRNN
- PointNet++ TensorFlow implementation: https://github.com/charlesq34/pointnet2
- Dynamic Graph CNN for Learning on Point Clouds https://github.com/WangYueFt/dgcnn
- Temporal Interpolation of Dynamic Point Clouds using Convolutional Neural Networks https://github.com/jelmr/pc_temporal_interpolation