A Pytorch implementation of the paper: Pixel2point: 3D Object Reconstruction From a Single Image Using CNN and Initial Sphere

• A. J. Afifi, J. Magnusson, T. A. Soomro, and O. Hellwich, “Pixel2point: 3D Object Reconstruction From a Single Image Using CNN and Initial Sphere, ” IEEE Access, vol. 9, pp. 110–121, 2021, doi: 10.1109/ACCESS.2020.3046951.

Input 2D Image	Ground Truth Point Cloud	Output Point Cloud

Environment

conda env create -f ./environment.yml
conda activate pixel2point

The code has been tested on Ubuntu 20.04, Python 3.9.13, Pytorch 1.12.1, Pytorch3D 0.7.0, CUDA 11.7

EMD Loss Function

The Earth Mover's Distance of point clouds is from: Colin97/MSN-Point-Cloud-Completion: Morphing and Sampling Network for Dense Point Cloud Completion (AAAI2020)

Dataset

• Shapenet Point cloud (shapenetcorev2_hdf5_2048.zip, 0.98G): antao97/PointCloudDatasets
• Shapenet renderer (image.tar, 30G): Xharlie/ShapenetRender_more_variation

required to modify the dataset path in the ./lib/settings.py

Training

python main.py

Testing

python test.py

required to modify the model path in ./lib/settings.py

Notes

• 2022.09.18

  • Dataset: Not tested on Pix3D
  • Model: Incomplete Fully Connected Layer because CUDA out of memory

• 2022.09.23

  • The "pytorch3d" chamfer distance on the CNN layer is non-deterministic at backward propagation on the GPU.
  • This repo uses the "pytorch" function to calculate the chamfer distance to make the loss function deterministic on the CNN layers (Increase execution time by 50%).