Home

Awesome

<h1 align="center"> ESFP: Event-based Shape from Polarization (CVPR 2023) </h1> <br>

This repository contains the code and download links to our dataset for our work on "Event-based Shape-from-Polarization", CVPR 2023 by Manasi Muglikar, Leonard Bauersfeld, Diederik Moeys, and Davide Scaramuzza.

<h2 align="center">

Project Page | Paper | Video | Dataset

</h2>

Event-based Shape from Polarization

Citation

If you use this code in an academic context, please cite the following work:

@InProceedings{Muglikar23CVPR,
  author = {Manasi Muglikar and Leonard Bauersfeld and Diederik Moeys and Davide Scaramuzza},
  title = {Event-based Shape from Polarization},
  booktitle = {IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR)},
  month = {Jun},
  year = {2023}
}

Installation

Install metavision from here

conda create -y -n esfp
conda activate esfp
conda install -y -c anaconda numpy scipy
conda install -y -c conda-forge opencv tqdm matplotlib pybind11 h5py blosc-hdf5-plugin
pip install --no-cache-dir -r training_code/requirements.txt

Dataset

We present the first large scale dataset consisting of several objects with different textures and shapes, and featuring multiple illumination and scene depths, for a total of 100 synthetic and 90 real scenes.

Download the dataset from here

To download the mistuba dataset and real dataset use the following links respectively:

wget https://download.ifi.uzh.ch/rpg/ESfP/mitsuba_dataset.zip
wget https://download.ifi.uzh.ch/rpg/ESfP/realworld_dataset.zip

Train

To train the network to predict surface normals, use the following training scripts: bash training_code/scripts/train_events_esfp_syn.sh