SPARC: Sparse Render-and-Compare for CAD model alignment from a single RGB image.

Official implementation of the paper

SPARC: Sparse Render-and-Compare for CAD model alignment from a single RGB image
British Machine Vision Conference 2022
Florian Langer, Gwangbin Bae, Ignas Budvytis, Roberto Cipolla
arXiv Project Page

Installation Requirements

• PyTorch
• PyTorch3D

We recommend installing via conda.

conda create -n sparc python=3.9
conda activate sparc
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
conda install pytorch3d -c pytorch3d

After installing the packages above install additional dependencies by

pip install -r requirements.txt

To install this repo

git clone https://github.com/florianlanger/SPARC
cd SPARC && pip install -e .

Data

Download data from this link: https://e1.pcloud.link/publink/show?code=XZ0oOQZ1nuA57shrQpAa7E2aGo8zz9AD5zX

• sparc_release
  • data
    • data_3d
    • data_scannet
    • train
    • val
  • main_experiment
    • results
      • per_frame_predictions.json
      • raw_results.csv
      • results_scannet.txt
    • visualisation
    • config.json
    • network.pth

The folder data contains all necessary data for training and evaluating SPARC on ScanNet25k. We also directly release our results. The results folder contains per frame predictions as well as the predictions transformed into ScanNet world coodinates raw_results.csv. The accuracies we obtain are provided in results_scannet.txt. We also visualise predictions for all images in visualisation.

Training/Evaluating

To train open the config file in the SPARC code and replace the tags ["general"]["output_dir"] and ["general"]["dataset_dir"] with the intended output dir path and the path to the downloaded and unzipped SPARC data. For trainig run python main.py. For evaluating run bash eval.sh. This will evaluate the provided model by first selecting one of four rotation initialisations for each image and then iteratively improving the pose for the best initialisation.

Citations

If you find our work helpful for your research please consider citing the following publication:

@inproceedings{sparc,
               author = {Langer, F. and Bae, G. and Budvytis, I. and Cipolla, R.},
               title = {SPARC: Sparse Render-and-Compare for CAD model alignment in a single RGB image},
               booktitle = {Proc. British Machine Vision Conference},
               month = {November},
               year = {2022},
               address={London}
}