HelloDyson/Sketch2Voxel

Final project repo of 6.869 Computer Vision MIT - 3D shape reconstruction from human sketch

Pixel2Voxel (pytorch)

Final project repo of 6.869 Computer Vision MIT

Original 3D-R2N2 with Theano: https://github.com/chrischoy/3D-R2N2

Original CycleGAN/Pix2pix page: https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix

Installation

This code tests on Ubuntu 18.04 with CUDA 9.0 but it should work on Windows/ Mac /Linux os

git clone this repository
cd Pytorch-3dr2n2
pip install -r pix2pix/requirements.txt

install PyTorch in https://pytorch.org/get-started/locally/

Dataset

Download the following dataset of 3D-R2N2 and save in data

ShapeNet rendered images http://cvgl.stanford.edu/data2/ShapeNetRendering.tgz

ShapeNet voxelized models http://cvgl.stanford.edu/data2/ShapeNetVox32.tgz

Download CycleGan/pix2pix dataset

cd pix2pix
bash ./scripts/download_pix2pix_model.sh edge2shoes_pretrained
bash ./datasets/download_pix2pix_dataset.sh edge2shoes

Test 3D reconstruction

open "3D-R2N2 3D Reconstruction.ipynb" and run all

Process data into pytorch dataloader format

open "Data preparing.ipynb" and run all

Train on PyTorch dataloader and model

open "Train on Pytorch.ipynb" and run all