Skip to content

TimoStoff/event_cnn_minimal

master
Switch branches/tags
Code

event_cnn_minimal

Minimal code for running inference on models trained for Reducing the Sim-to-Real Gap for Event Cameras, ECCV'20.

Running with Anaconda

cuda_version=10.1

conda create -y -n event_cnn python=3.7
conda activate event_cnn
conda install -y pytorch torchvision cudatoolkit=$cuda_version -c pytorch
conda install -y -c conda-forge opencv
conda install -y -c conda-forge tqdm
conda install -y -c anaconda h5py 
conda install -y -c intel pandas
conda install -y -c anaconda scikit-image
pip install thop --user

As a further prerequisite, you will need to have ROS installed on your system. Make sure not to source your ROS and Conda envs at the same time, as they conflict.

Usage

Clone this repo and submodules:

git clone -b inference git@github.com:TimoStoff/event_cnn_minimal.git --recursive
cd event_cnn_minimal/events_contrast_maximization/
git checkout master
cd ..

Conversion to HDF5

This code processes the events in HDF5 format. To convert the rosbags to this format, open a new terminal and source a ROS workspace.

source /opt/ros/kinetic/setup.bash
python events_contrast_maximization/tools/rosbag_to_h5.py <path/to/rosbag/or/dir/with/rosbags> --output_dir <path/to/save_h5_events> --event_topic <event_topic> --image_topic <image_topic>

As an example, using slider_depth from "The event camera dataset and simulator":

wget http://rpg.ifi.uzh.ch/datasets/davis/slider_depth.bag -O /tmp/slider_depth.bag
source /opt/ros/kinetic/setup.bash
python events_contrast_maximization/tools/rosbag_to_h5.py /tmp/slider_depth.bag --output_dir /tmp/h5_events --event_topic /dvs/events --image_topic /dvs/image_raw

If you have access to events from a color event camera, you need to set image_topic to the topic containing events and a flag --is_color. For example, using carpet_simple.bag:

python events_contrast_maximization/tools/rosbag_to_h5.py /tmp/simple_carpet.bag --image_topic /dvs/image_color --is_color

Inference

Download the pretrained models from here, into event_cnn_minimal.

To estimate reconstruction:

python inference.py --checkpoint_path <path/to/model.pth> --device 0 --h5_file_path </path/to/events.h5> --output_folder </path/to/output/dir>

For example:

python inference.py --checkpoint_path pretrained/reconstruction/reconstruction_model.pth --device 0 --h5_file_path /tmp/h5_events/slider_depth.h5 --output_folder /tmp/reconstruction

To estimate flow:

python inference.py --checkpoint_path <path/to/model.pth> --device 0 --h5_file_path </path/to/events.h5> --output_folder </path/to/output/dir> --is_flow

For example:

python inference.py --checkpoint_path pretrained/flow/flow_model.pth --device 0 --h5_file_path /tmp/h5_events/slider_depth.h5 --output_folder /tmp/flow --is_flow

Flow is saved as both a png showing HSV color as slow vectors and as npy files. Should look something like this: Reconstruction Flow Color

We provide some of our more recent models for download. These models are prefixed with update (ie: update_flow_model.pth). These models have not necessarily been quantitatively evaluated and are not necessarily better than the models reported on in the paper Reducing the Sim-to-Real Gap for Event Cameras. If you wish to run these models, the flag --update is necessary, as the voxels are formed in a slightly different way in the updated models. Again, the models from the paper are the ones without the prefix 'update'.

Training dataset

You will need to generate the training dataset yourself, using ESIM. To find out how, please see the training data generator repo.

Training

To train a model, you need to create a config file (see config/config.json for an example). In this file, you need to set what model you would like to use (you can choose from several models from the literature such as EVFlowNet etc, see the files in model/ for more. You also need to set the training parameters, the training data, the validation data and the output directory. To train the flow network and reconstruction network from the paper, see config/flow.json and config/reconstruction.json respectively. You can then start the training by invoking

python train.py --config path/to/config

If you have a model that would like to keep training from, you can use

python train.py --config path/to/config --resume /path/to/model.pth

Citations

Please cite the following if you decide to use this code in an academic context:

@Article{stoffregen2020eccv,
    title={Reducing the Sim-to-Real Gap for Event Cameras},
    author={Timo Stoffregen and Cedric Scheerlinck and Davide Scaramuzza and Tom Drummond and Nick Barnes and Lindsay Kleeman and Robert Mahony},
    journal=eccv,
    year=2020,
    month=aug
}