The code accompanying our ECCV 2020 publication and dataset, EVE.

• Authors: Seonwook Park, Emre Aksan, Xucong Zhang, and Otmar Hilliges
• Project page: https://ait.ethz.ch/eve
• Codalab (test set evaluation and public leaderboard): https://competitions.codalab.org/competitions/28954

Preferably, setup a Docker image or virtual environment (virtualenvwrapper is recommended) for this repository. Please note that we have tested this code-base in the following environments:

• Ubuntu 18.04 / A Linux-based cluster system (CentOS 7.8)
• Python 3.6 / Python 3.7
• PyTorch 1.5.1

Clone this repository somewhere with:

git clone [email protected]:swook/EVE
cd EVE/

Then from the base directory of this repository, install all dependencies with:

pip install -r requirements.txt

Please note the PyTorch official installation guide for setting up the torch and torchvision packages on your specific system.

You will also need to setup ffmpeg for video decoding. On Linux, we recommend installing distribution-specific packages (usually named ffmpeg). If necessary, check out the official download page or compilation instructions.

All available configuration parameters are defined in src/core/config_default.py.

In order to override the default values, one can do:

1. Pass the parameter via a command-line parameter to train.py or inference.py. Note that in this case, replace all _ characters with -. E.g. the config. parameter refine_net_enabled becomes --refine-net-enabled 1. Note that boolean parameters can be passed in via either 0/no/false or 1/yes/true.
2. Create a JSON file such as src/configs/eye_net.json or src/configs/refine_net.json.

The order of application are:

1. Default parameters
2. JSON-provided parameters, in order of JSON file declaration. For instance, in the command python train.py config1.json config2.json, config2.json overrides config1.json entries should there be any overlap.
3. CLI-provided parameters.

This framework implements an automatic logging code of all parameters, loss terms, and metrics to a Google Sheets document. This is done by the gspread library. To enable this possibility, follow these instructions:

1. Follow the instructions at https://gspread.readthedocs.io/en/latest/oauth2.html#for-end-users-using-oauth-client-id
2. Set --gsheet-secrets-json-file to a path to the credentials JSON file, and set --gsheet-workbook-key to the document key. This key is the part after https://docs.google.com/spreadsheets/d/ and before any query or hash parameters.

An example config JSON file can be found at src/configs/sample_gsheet.json.

To train a model, simply run python train.py from src/ with the appropriate configuration changes that are desired (see "Configuration file system" above).

Note, that in order to resume the training of an existing model you must provide the path to the output folder via the --resume-from argument.

Also, at every fresh run of train.py, a unique identifier is generated to produce a unique output folder in outputs/EVE/. Hence, it is recommended to use the Google Sheets logging feature (see "Automatic logging to Google Sheets") to keep track of your models.

The single-sample inference script at src/inference.py takes in the same arguments as train.py but expects two arguments in particular:

• --input-path is the path to a basler.mp4 or webcam_l.mp4 or webcam_c.mp4 or webcam_r.mp4 that exists in the EVE dataset.
• --output-path is a path to a desired output location (ending in .mp4).

This script works for both training, validation, and test samples and shows the reference point-of-gaze ground-truth when available.

If using this code-base and/or the EVE dataset in your research, please cite the following publication:

@inproceedings{Park2020ECCV,
  author    = {Seonwook Park and Emre Aksan and Xucong Zhang and Otmar Hilliges},
  title     = {Towards End-to-end Video-based Eye-Tracking},
  year      = {2020},
  booktitle = {European Conference on Computer Vision (ECCV)}
}

Q: How do I use this code for screen-based eye tracking?

A: This code does not offer actual eye tracking. Rather, it concerns the benchmarking of the video-based gaze estimation methods outlined in the original paper. Extending this code to support an easy-to-use software for screen-based eye tracking is somewhat non-trivial, due to requirements on camera calibration (intrinsics, extrinsics), and an efficient pipeline for accurate and stable real-time eye or face patch extraction. Thus, we consider this to be beyond the scope of this code repository.

Q: Where are the test set labels?

A: Our public evaluation server and leaderboard are hosted by Codalab at https://competitions.codalab.org/competitions/28954. This allows for evaluations on our test set to be consistent and reliable, and encourage competition in the field of video-based gaze estimation. Please note that the performance reported by Codalab is not strictly speaking comparable to the original paper's results, as we only perform evaluation on a large subset of the full test set. We recommend acquiring the updated performance figures from the leaderboard.