deepSDF-encoder

This repo contains a complete implementation of a encoder-decoder model that takes in point cloud observations and represents fine-grained 3D geometry as implicit signed distance field. It is adapted from DI-Fusion and DeepSDF.

Installation

1. System Dependencies

• Ubuntu >= 18.04
• cuda >= 10.2
• cudnn 7
• Python 3.7

2. Python Packages

We recommend installing python dependencies in a virtual environment (python 3.7):

pip3 install -r requirements.txt

3. Dependencies for Data Generation

To generate training data yourself, please install the following dependencies:

• CLI11
• Eigen-3.3.9
  • Use other versions may cause issues
• Pangolin
  • Make sure you use the stable release version
  • May need to install mpark/variant
• Flann
  • Make sure cuda toolkit is correctly installed and configured
  • Enable BUILD_CUDA_LIB before running 'cmake ..'

The general process to install the packages:

cd <your-lib-folder>
# 1. Download the package or use git clone, for example:
git clone https://github.com/CLIUtils/CLI11.git
# 2. Create build folder under package path, for example:
cd CLI11 && mkdir build && cd build
# 3. Run cmake
cmake ..
# 4. Run make and install to the system
make
make install

Data Generation

First build the cuda-based mesh sampler:

cd <this-repo>/sampler_cuda
mkdir build && cd build
cmake ..
make -j

Then run:

# Enable headless rendering
export PANGOLIN_WINDOW_URI=headless://
python data_generator_sdf.py configs/data-cutting-shape.yaml --nproc 8

Note that you need to specify the input and output path in configs/data-cutting-shape.yaml.

Training

python network_trainer_sdf.py configs/train-sdf.yaml [--options]

Please refer to configs/train-sdf.yaml for different running options. The code supports parallel training.

You can also use tensorboard to monitor the training process:

tensorboard --logdir di-checkpoints/train_cutting/tensorboard