Home

Awesome

davis-EDI-rs

A fast, Rust-based, open-source implementation of the paper "Bringing a Blurry Frame Alive at High Frame-Rate with an Event Camera" (2019) by Pan et al.

About

This project aims to elucidate the paper above and move towards real-time software systems which take advantage of synthesized event and frame data. The original paper's code is largely obfuscated, making it impossible to make improvements. Furthermore, the original code base is written in MATLAB, making it relatively slow, and it uses a converted MATLAB data format for the DAVIS files.

This implementation operates on .aedat4 files generated directly by an iniVation DAVIS camera and a live feed from a DAVIS camera. If we're going to move towards practical systems with event cameras, we need practical processing! I've included a sample file recorded with my camera.

I did my best to understand and implement the mathematics presented in the original paper, while taking a few liberties for the sake of speed gains. At present, this project only implements the Event-based Double Integral from the 2019 paper, not the multiple Event-based Double Integral from the 2020 paper.

Usage

Since there are a lot of command-line arguments available, my preferred method of serving them is in a file. Refer to Args.toml for an example. You can give it a run by cloning this repository and executing:

cargo run --release -- --args-filename "./Args.toml"

Deblur from an .aedat4 file

You can deblur a pre-existing file by providing the directory of the file in --base-path, the file name in --event-filename-0, and --mode to "file".

Deblur from a live camera feed

You can also deblur the data coming straight from a camera, in real time! I've provided the config file for iniVation's DV software which lets you publish the APS frames and event packets to two Unix sockets. For this approach, you should set --base-path to "/tmp", --event-filename-0 to the name of the events socket, --event-filename-1 to the name of the frames socket, and --mode to "socket". This should work pretty much the same way for a TCP connection, but additional configuration may be required.

Other parameters

--output-fps: The reconstruction output frame rate. Increasing this parameter has a marginal effect on the processing speed.

--show-display: Whether or not to show a live view of the reconstruction, using OpenCV display windows.

--write-video: If true, writes the reconstructed frames to an .avi file.

--optimize-c: If true, will dynamically choose the optimal contrast threshold for deblurring each frame. Causes a significant slow down, especially for higher frame-rate inputs, and can make the reconstruction perform slightly less than real time. If false, then the --start-c value provided will be the contrast threshold used for deblurring all frames.

--optimize-controller: If true, will attempt to maintain real-time reconstruction performance. The controller dynamically toggles whether contrast threshold optimization is performed (unless --optimize-c is false), and adjusts the reconstruction frame rate. If false, will maintain a constant reconstruction frame rate, but may fall behind real-time performance. The reconstructed video files will be much smoother with this disabled. If the scene dynamics won't change much, and you have the ability to dial in settings ahead of time, it's best to keep this disabled and find (through trail and error) the optimal --output-fps value which maintains good performance.

To-do list

There are some major things left before I can start implementing mEDI. Any assistance from the community would be greatly appreciated

Requirements

Compatibility

Only tested on Ubuntu 20.04, using dv-gui 1.6 for piping live camera data. The default Args.toml won't work on Windows because it uses Unix sockets, but it could probably work through files or TCP. Builds on both Ubuntu and Windows as demonstrated through CI (macOS yet to be tested).

Release notes

v0.2.3, 2023-09-26

v0.2.0, 2023-03-08

Cite this work

If you write a paper which references this software, we ask that you reference the following papers on which it is based. Citations are given in the BibTeX format.

An Asynchronous Intensity Representation for Framed and Event Video Sources

(Main repository)

@inproceedings{10.1145/3587819.3590969,
author = {Freeman, Andrew C. and Singh, Montek and Mayer-Patel, Ketan},
title = {An Asynchronous Intensity Representation for Framed and Event Video Sources},
year = {2023},
isbn = {979-8-4007-0148-1/23/06},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3587819.3590969},
doi = {10.1145/3587819.3590969},
booktitle = {Proceedings of the 14th ACM Multimedia Systems Conference},
pages = {1–12},
numpages = {12},
location = {Vancouver, BC, Canada},
series = {MMSys '23}
}

The ADΔER Framework: Tools for Event Video Representations

@inproceedings{Freeman23-0,
  title = {The ADΔER Framework: Tools for Event Video Representations},
  author = {Andrew C. Freeman},
  year = {2023},
  doi = {10.1145/3587819.3593028},
  url = {https://doi.org/10.1145/3587819.3593028},
  researchr = {https://researchr.org/publication/Freeman23-0},
  cites = {0},
  citedby = {0},
  pages = {343-347},
  booktitle = {Proceedings of the 14th Conference on ACM Multimedia Systems, MMSys 2023, Vancouver, BC, Canada, June 7-10, 2023},
  publisher = {ACM},
}