Code examples for Online learning of Continuous Signed Distance Fields Using Piecewise Polynomials.

This repository started as a fork of the RDF codebase.

Tested with Python 3.10.12 on Ubuntu 22.04 and macOS 14.4.1.

Python dependencies are listed in requirements.txt. Open3D additionally requires a working C++ compiler.

(Optional) For GPU acceleration, a compatible version of CUDA is required (tested with CUDA 11.5).

Setup a conda new environment:

conda create --name ppSDF python=3.10.12

Activate the environment:

conda activate ppSDF

Install the required packages:

pip install -r requirements.txt

The desired objects from the YCB dataset (Calli et al.) can be downloaded by running

python ycb_downloader.py

Objects can be selected by commenting/uncommenting elements in the objects_to_download list inside the script.

To run the example script with the desired arguments:

• n_seg - number of segments per input dimension
• qd, qn, qt - cost coefficients for incremental learning
• sigma - measurement noise
• batch_size - batch size for the incremental updates
• device - device to run on (e.g., cuda or cpu)
• save - if True, saves the approximated SDF
• object - which object to load
• n_data - number of training sample points
• cut_x, cut_y, cut_z - select axis to cut for visualization

For example, to approximate the '035_power_drill' object with 5 segments per input dimension and 1000 training samples:

python reconstruct.py --object=035_power_drill --n_seg=5 --n_data=1000

To run with default arguments:

python reconstruct.py

Maintained by Ante MARIĆ and licensed under the GPL-3.0 License.

Copyright (c) 2024 Idiap Research Institute [email protected]