Image to Icosahedral Projection for SO(3) Object Reasoning from Single-View Images

Paper | Project Page

---

This is the code for "Image to Icosahedral Projection for SO(3) Object Reasoning from Single-View Images" presented at NeurReps Workshop at NeurIPS 2022 and to appear in PMLR Volume on Symmetry and Geometry.

Installation

The code was tested using Python 3.8, and the neural networks are instantiated with PyTorch.

pip install -r requirements.txt

Downloading and Preparing Datasets

Run the following scripts to download the object files, then render them into images. Note, you will need an account at shapenet.org to download the shapenet object files.

cd datasets
python render_objects.py --dataset=modelnet40 --mode=depth --num_views=60
python render_objects.py --dataset=modelnet40 --mode=gray --num_views=60
python render_objects.py --dataset=shapenet55 --mode=rgb --num_views=60
cd ..

Rendering can take several hours for all object classes, you can selectively render object classes using --objects arg.

Training Orientation Prediction

python run.py --task=orientation --dataset=modelnet40 --objects=select --mode=depth

Training Shape Classification

python run.py --task=classification --dataset=modelnet40 --objects=all --mode=depth