This is the PyTorch implementation of the ICCV 2021 paper Sketch2Mesh. We provide pre-trained networks and code for demonstrating our global differentiable refinement procedure of 3D shapes from 2D sketches.

The below instructions describe how to:

1. setup the python environment
2. download data
3. download pre-trained networks
4. launch refinement
5. read metrics

Set up a conda environment with the right packages using:

conda env create -f conda_env.yml
conda activate sketch2mesh

We provide our test set in the form of an archive containing sketches (~8MB) and the associated meshes (~530MB). The latter is a subset of ShapeNet meshes that were pre-processed by DISN authors.

To download both sketches and meshes directly at the right location using gdown (already installed if you followed the above setup with conda), from the root folder of the cloned repo use:

cd dataset
gdown https://drive.google.com/uc?id=1Zwp5MdvHY13zjF5KndueBpSBsqW_9Ip1
gdown https://drive.google.com/uc?id=1iAr12e3cqribB7jDGogLToxtzRNZP82M
unzip Sketches.zip
unzip Meshes.zip
rm Sketches.zip
rm Meshes.zip
cd ..

Pre-trained encoder-decoder networks for sketches of cars and chairs (~35MB each) should be downloaded to the output directory using:

mkdir output
cd output
gdown https://drive.google.com/uc?id=1C09_0RMiG2on8rvEqo3z79GzDoGvI3I2
gdown https://drive.google.com/uc?id=1MEf4p-MaSVzL9v3i1GTMzJogM_0ciz6y
unzip cars.zip
unzip chairs.zip
rm cars.zip
rm chairs.zip
cd ..

These networks were trained on the Suggestive sketching style (see main paper).

Reconstruction and refinement from a collection of input sketches is done in reconstruct_sketch2mesh.py, with the following options:

• --experiment : relative path to either the chairs (output/chairs) or cars (output/cars) pre-trained network directory,
• --out_dir : output sub-directory where optimized meshes will be stored,
• --sketch_style : style of the input sketches, from [fd | suggestive | handdrawn (for cars only)]. See the main paper (Sec. 4.1) for more explanations.

For example, to launch reconstruction + refinement on hand drawn cars:

python reconstruct_sketch2mesh.py --experiment output/cars --out_dir sugg_to_hand --sketch_style handdrawn

We provide a very simplistic parallelized implementation, using additional options --n and --N. The list of test shapes is divided into N equal chunks, and the n-th launched process handles the n-th chunk. For example, launching 6 reconstruction threads in parallel for the above reconstruction is done with

(python reconstruct_sketch2mesh.py --experiment output/cars --out_dir sugg_to_hand --sketch_style handdrawn --N 6 --n 1 &
python reconstruct_sketch2mesh.py --experiment output/cars --out_dir sugg_to_hand --sketch_style handdrawn --N 6 --n 2 &
python reconstruct_sketch2mesh.py --experiment output/cars --out_dir sugg_to_hand --sketch_style handdrawn --N 6 --n 3 &
python reconstruct_sketch2mesh.py --experiment output/cars --out_dir sugg_to_hand --sketch_style handdrawn --N 6 --n 4 &
python reconstruct_sketch2mesh.py --experiment output/cars --out_dir sugg_to_hand --sketch_style handdrawn --N 6 --n 5 &
python reconstruct_sketch2mesh.py --experiment output/cars --out_dir sugg_to_hand --sketch_style handdrawn --N 6 --n 6)

Once the above has been performed, provide read_metrics.py with the path to reconstructed shapes to get 3D metrics before/after refinement. For example:

python read_metrics.py -d output/cars/Optimizations/latest/sugg_to_hand/ShapeNetV2/02958343/
> Across 113 shapes:
>  - Average initial 3D Chamfer: 6.835395413599249
>  - After refinement: 3.7756478764215666