A Differentiable Recurrent Surface for Asynchronous Event-Based Data

Code for the ECCV2020 paper "A Differentiable Recurrent Surface for Asynchronous Event-Based Data"
Authors: Marco Cannici, Marco Ciccone, Andrea Romanoni, Matteo Matteucci

Citing:

If you use Matrix-LSTM for research, please cite our accompanying ECCV2020 paper:

@InProceedings{Cannici_2020_ECCV,
    author = {Cannici, Marco and Ciccone, Marco and Romanoni, Andrea and Matteucci, Matteo},
    title = {A Differentiable Recurrent Surface for Asynchronous Event-Based Data},
    booktitle = {The European Conference on Computer Vision (ECCV)},
    month = {August},
    year = {2020}
}

Project Structure

The code is organized in two folders:

• classification/ containing PyTorch code for N-Cars and N-Caltech101 experiments
• opticalflow/ containing TensorFlow code for MVSEC experiments (code based on EV-FlowNet repository)

Note: the naming convention used within the code is not exactly the same as the one used in the paper. In particular, the groupByPixel operation is named group_rf_bounded in the code (i.e., group by receptive field, since it also supports receptive fields larger than 1x1), while the groupByTime operation is named intervals_to_batch.

Requirements

We provide a Dockerfile for both codebases in order to replicate the environments we used to run the paper experiments. In order to build and run the containers, the following packages are required:

• Docker CE - version 18.09.0 (build 4d60db4)
• NVIDIA Docker - version 2.0

If you have installed the latest version, you may need to modify the .sh files substituting:

• nvidia-docker run with docker run
• --runtime=nvidia with --gpus=all

You can verify which command works for you by running:

• (scripts default) nvidia-docker run -ti --rm --runtime=nvidia -t nvidia/cuda:10.0-cudnn7-devel-ubuntu16.04 nvidia-smi
• docker run -ti --rm --gpus=all -t nvidia/cuda:10.0-cudnn7-devel-ubuntu16.04 nvidia-smi

You should be able to see the output of nvidia-smi

Run Experiments

Details on how to run experiments are provided in separate README files contained in the classification and optical flow sub-folders:

• Classification Experiments
• Optical Flow Experiments

Note: using Docker is not mandatory, but it will allow you to automate the process of installing dependencies and building CUDA kernels, all within a safe environment that won't modify any of your previous installations. Please, read the Dockerfile and requirements.yml files contained inside the <classification or opticalflow>/docker/ subfolders if you want to perform a standard conda/pip installation (you just need to manually run all RUN commands).