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