/rpg_vid2e

Open source implementation of CVPR 2020 "Video to Events: Recycling Video Dataset for Event Cameras"

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Video to Events: Recycling Video Datasets for Event Cameras

Video to Events

This repository contains code that implements video to events conversion as described in Gehrig et al. CVPR'20 and the used dataset. The paper can be found here

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

Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, Davide Scaramuzza, "Video to Events: Recycling Video Datasets for Event Cameras", The Conference on Computer Vision and Pattern Recognition (CVPR), 2020

@InProceedings{Gehrig_2020_CVPR,
  author = {Daniel Gehrig and Mathias Gehrig and Javier Hidalgo-Carri\'o and Davide Scaramuzza},
  title = {Video to Events: Recycling Video Datasets for Event Cameras},
  booktitle = {{IEEE} Conf. Comput. Vis. Pattern Recog. (CVPR)},
  month = {June},
  year = {2020}
}

News

  • We now support frame interpolation done by FILM.
  • We release a web app and interactive demo which generates events and converts your webcam to events. Try it out here.
  • We now also release new python bindings for esim with GPU support. Details are here

Web App and Interactive Demo

Try out our the interactive demo and webcam support here.

Dataset

The synthetic N-Caltech101 dataset, as well as video sequences used for event conversion can be found here. For each sample of each class it contains events in the form class/image_%04d.npz and images in the form class/image_%05d/images/image_%05d.png, as well as the corresponding timestamps of the images in class/image_%04d/timestamps.txt.

Installation

Clone the repo recursively with submodules

git clone git@github.com:uzh-rpg/rpg_vid2e.git --recursive

Installation

First download the FILM checkpoint, and move it to the current root

    wget https://rpg.ifi.uzh.ch/data/VID2E/pretrained_models.zip -O /tmp/temp.zip
    unzip /tmp/temp.zip -d rpg_vid2e/
    rm -rf /tmp/temp.zip

make sure to install the following * Anaconda Python 3.9 * CUDA Toolkit 11.2.1 * cuDNN 8.1.0

conda create --name vid2e python=3.9
conda activate vid2e
pip install -r rpg_vid2e/requirements.txt
conda install -y -c conda-forge pybind11 matplotlib
conda install -y pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Build the python bindings for ESIM

pip install rpg_vid2e/esim_py/

Build the python bindings with GPU support with

pip install rpg_vid2e/esim_torch/

Adaptive Upsampling

This package provides code for adaptive upsampling with frame interpolation based on Super-SloMo

Consult the README for detailed instructions and examples.

esim_py

This package exposes python bindings for ESIM which can be used within a training loop.

For detailed instructions and example consult the README

esim_torch

This package exposes python bindings for ESIM with GPU support.

For detailed instructions and example consult the README

Example

To run an example, first upsample the example videos

device=cpu
# device=cuda:0
python upsampling/upsample.py --input_dir=example/original --output_dir=example/upsampled --device=$device

This will generate upsampling/upsampled with in the example/upsampled folder. To generate events, use

python esim_torch/generate_events.py --input_dir=example/upsampled \
                                     --output_dir=example/events \
                                     --contrast_threshold_neg=0.2 \
                                     --contrast_threshold_pos=0.2 \
                                     --refractory_period_ns=0