This is a code repo for Learning to Super Resolve Intensity Images from Events (CVPR 2020 - Oral)
Mohammad Mostafavi, Jonghyun Choi and Kuk-Jin Yoon (Corresponding author)

Our extended and upgraded version produces highly consistent videos, and includes further details and experiments E2SRI: Learning to Super-Resolve Intensity Images from Events - TPAMI 2021

If you use any of this code, please cite both following publications:

@article{mostafaviisfahani2021e2sri,
  title     = {E2SRI: Learning to Super-Resolve Intensity Images from Events},
  author    = {Mostafaviisfahani, Sayed Mohammad and Nam, Yeongwoo and Choi, Jonghyun and Yoon, Kuk-Jin},
  journal   = {IEEE Transactions on Pattern Analysis \& Machine Intelligence},
  number    = {01},
  pages     = {1--1},
  year      = {2021},
  publisher = {IEEE Computer Society}
}

@article{mostafavi2020e2sri,
  author  = {Mostafavi I., S. Mohammad and Choi, Jonghyun and Yoon, Kuk-Jin},
  title   = {Learning to Super Resolve Intensity Images from Events},
  journal = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  month   = {June},
  year    = {2020},
  pages   = {2768-2786}
}

• Mohammad Mostafavi
• Yeong-oo Nam

• Make your own environment

python -m venv ./e2sri
source e2sri/bin/activate

• Install the requirements

cd e2sri
pip install -r requirements.txt

• Unizp pyflow

cd src
unzip pyflow.zip
cd pyflow
python3 setup.py build_ext -i

• Download the linked material below

  • Sample pretrained weight (2x_7s.pth) for 2x scale (2x width and 2x height) and 7S sequences of stacks.
  • Sample dataset for training and testing (datasets.zip).

• Unzip the dataset.zip file and put the pth weight file in the main folder

unzip dataset.zip
cd src

• Run inference:

python test.py --data_dir ../dataset/slider_depth --checkpoint_path ../save_dir/2x_7s.pth --save_dir ../save_dir

Note that our code with the given weights (7S) consumes ~ 4753MiB GPU memory at inference.

From this sample event stack, you should produce a similar (resized) result as:

• Run training:

python3 train.py --config_path ./configs/2x_3.yaml --data_dir ../dataset/Gray_5K_7s_tiny --save_dir ../save_dir

We provided a sample sequence (slider_depth.zip) made from the rosbags of the Event Camera Dataset and Simulator. The rosbag (bag) is a file format in ROS (Robot Operating System) for storing ROS message data. You can make other sequences using the given matlab m-file (/e2sri/stacking/make_stacks.m). The matlab code depends on matlab_rosbag which is included in the stacking folder and needs to be unzipped.

Note: The output image quality relies on "events_per_stack" and "stack_shift". We used "events_per_stack"=5000, however we did not rely on "stack_shift" as we synchronized with APS frames instead. The APS synchronized stacking when this 5000 setting should be kept will be released with the training code together.

A list of publicly available event datasets for testing:

• Bardow et al., CVPR'16
• The Event Camera Dataset and Simulator
• Multi Vehicle Stereo Event Camera Dataset, RAL'18
• Scherlinck et al., ACCV'18
• High Speed and HDR Dataset
• Color event sequences from the CED dataset Scheerlinck et al., CVPR'18

• Stereo Depth from Events Cameras: Concentrate and Focus on the Future + Code - CVPR 2022 (TBU)

• Event-Intensity Stereo: Estimating Depth by the Best of Both Worlds - Openaccess ICCV 2021 (PDF)

• E2SRI: Learning to Super Resolve Intensity Images from Events - TPAMI 2021 (Link)

• Learning to Reconstruct HDR Images from Events, with Applications to Depth and Flow Prediction - IJCV 2021

• Learning to Super Resolve Intensity Images from Events - Openaccess CVPR 2020 (PDF)

• Event-Based High Dynamic Range Image and Very High Frame Rate Video Generation Using Conditional Generative Adversarial Networks - Openaccess CVPR 2019 (PDF)

MIT license.