Awesome

TimeLens: Event-based Video Frame Interpolation

This repository is about the High Speed Event and RGB (HS-ERGB) dataset, used in the 2021 CVPR paper TimeLens: Event-based Video Frame Interpolation by Stepan Tulyakov*, Daniel Gehrig*, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza.

For more information, visit our project page.

Citation

A pdf of the paper is available here. If you use this dataset, please cite this publication as follows:

@Article{Tulyakov21CVPR,
  author        = {Stepan Tulyakov and Daniel Gehrig and Stamatios Georgoulis and Julius Erbach and Mathias Gehrig and Yuanyou Li and
                  Davide Scaramuzza},
  title         = {{TimeLens}: Event-based Video Frame Interpolation},
  journal       = "IEEE Conference on Computer Vision and Pattern Recognition",
  year          = 2021,
}

Google Colab

A Google Colab notebook is now available here. You can upsample your own video and events from you gdrive.

Gallery

For more examples, visit our project page.

Installation

Install the dependencies with

cuda_version=10.2
conda create -y -n timelens python=3.7
conda activate timelens
conda install -y pytorch torchvision cudatoolkit=$cuda_version -c pytorch
conda install -y -c conda-forge opencv scipy tqdm click

Test TimeLens

First start by cloning this repo into a new folder

mkdir ~/timelens/
cd ~/timelens
git clone https://github.com/uzh-rpg/rpg_timelens

Then download the checkpoint and data to the repo

cd rpg_timelens
wget http://download.ifi.uzh.ch/rpg/web/data/timelens/data2/checkpoint.bin
wget http://download.ifi.uzh.ch/rpg/web/data/timelens/data2/example_github.zip
unzip example_github.zip 
rm -rf example_github.zip

Running Timelens

To run timelens simply call

skip=0
insert=7
python -m timelens.run_timelens checkpoint.bin example/events example/images example/output $skip $insert

This will generate the output in example/output. The first four variables are the checkpoint file, image folder and event folder and output folder respectively. The variables skip and insert determine the number of skipped vs. inserted frames, i.e. to generate a video with an 8 higher framerate, 7 frames need to be inserted, and 0 skipped.

The resulting images can be converted to a video with

ffmpeg -i example/output/%06d.png timelens.mp4

the resulting video is timelens.mp4.

Dataset

Download the dataset from our project page. The dataset structure is as follows

.
├── close
│   └── test
│       ├── baloon_popping
│       │   ├── events_aligned
│       │   └── images_corrected
│       ├── candle
│       │   ├── events_aligned
│       │   └── images_corrected
│       ...
│
└── far
    └── test
        ├── bridge_lake_01
        │   ├── events_aligned
        │   └── images_corrected
        ├── bridge_lake_03
        │   ├── events_aligned
        │   └── images_corrected
        ...

Each events_aligned folder contains events files with template filename %06d.npz, and images_corrected contains image files with template filename %06d.png. In events_aligned each event file with index n contains events between images with index n-1 and n, i.e. event file 000001.npz contains events between images 000000.png and 000001.png. Each event file contains keys for the x,y,t, and p event component. Note that x and y need to be divided by 32 before use. This is because they actually correspond to remapped events, which have floating point coordinates.

Moreover, images_corrected also contains timestamp.txt where image timestamps are stored. Note that in some folders there are more image files than event files. However, the image stamps in timestamp.txt should match with the event files and the additional images can be ignored.

For a quick test download the dataset to a folder using the link sent by email.

wget download_link.zip -O /tmp/dataset.zip
unzip /tmp/dataset.zip -d hsergb/

And run the test

python test_loader.py --dataset_root hsergb/ \ 
                      --dataset_type close \ 
                      --sequence spinning_umbrella \ 
                      --sample_index 400
                                          

This should open a window visualizing aligned events with a single image.