About the challenge
Immersive AR/VR demands unprecedented eye tracking performance. Eye tracking must be precise, accurate, and work all the time, for every person, in any environment. While advancements in deep learning have yielded successes in domains with similar challenges, the real time requirement and platform power limitations put serious memory and compute constraints on any ML-based solution. Additionally, a robust and efficient ML solution that is insensitive to environmental factors requires large amounts of highly accurate ground-truth training data from thousands of users in challenging conditions. Unfortunately, capturing accurate eye-gaze data in these environments requires a highly sophisticated and costly setup, and even with this setup, accuracy is limited by user fixation ability and cooperation. These issues place practical limitations on the amount and quality of training data that can be collected.
In the absence of accurate gaze labels, we propose to advance the state of the art by carefully designing two challenges that combine human annotation of eye features with unlabeled data. These challenges focus on deeper understanding of the distribution underlying human eye state. We invite ML and CV researchers for participation.
Track-1 Semantic Segmentation challenge: Many eye-tracking solutions require accurate estimation of eye-features in 2d images, typically per-pixel segmentation of the key eye regions: the sclera, the iris, the pupil, and everything else (background). Though eye segmentation solutions have been demonstrated [1,2], the ideal solution must be accurate, robust, and extremely power efficient. Therefore in this challenge, we evaluate both the accuracy of the model and approximate complexity using the model size as explained in Section 1.2.a of this document. This challenge encourages:
- Accurate and generalizable semantic segmentation solutions under memory-and-compute constraints.
- Learning and respecting the natural representation, including the geometry, of human eyes in the semi-supervised OpenEDS dataset
- Balancing accuracy and inference complexity while designing the semantic segmentation models
- Leveraging synthetic data generation if and where appropriate (ex. UnityEyes, NVGaze)
Track-2 Synthetic Eye Generation challenge: Most learning-based system achieve better performance and generalizability with more data. However, as explained earlier, capturing accurate real-world eye-gaze data at the scale required for training from image to gaze directly is difficult. For this challenge, we instead focus on generating realistic eye-data. Specifically, we propose a novel image-synthesis problem that aims to capture subject-specific signals from a few eye-images of an individual and generate realistic eye-images for the same individual under different eye-states (gaze direction, camera position, eye openness etc.). For this challenge, we posit that substantial information about the eye state is encoded in the feature segmentation masks similar to those derived in the Semantic Segmentation challenge. This task puts the current GAN and VAE based image synthesis models to an unprecedented test wherein exact pixel-level matching is required instead of high-level perceptual differences. This task encourages-
- Formulating image-synthesis models, such as GAN/VAEs, to extrapolate the information contained in a few images of subject’s eye with the given segmentation mask to new images.
- Incorporation of geometric modeling in GAN/VAE frameworks for constraining the image-synthesis with the physics of the human eye-system. For example, iris and pupil should be very close to an ellipse in their 2D image projection.
- Capturing subject-specific signals from a few images and extrapolating it into new gaze-directions, camera viewpoints and/or eyelid states.
Note: The task requires generating a realistic eye image, I, from a given semantic segmentation mask, M, of the same person, P. We have provided three JSON files to map the eye images to different subjects in the Train, Val, and Test datasets. Please use the provided “image/mask to identity” map and generate realistic eye images for a given segmentation mask of the same subject.
For this task, you can use all the training image/masks and try to achieve the best performance on the given test set of semantic segmentation masks. The metric used for measuring the performance is L2 distance from the original eye image.
OpenEDS is a data set of eye images captured using a virtual-reality HMD with two synchronized eye-facing cameras at a frame rate of 200 Hz under controlled illumination. The paper describing OpenEDS is available here.
This dataset is composed of:
- Semantic segmentation data set collected with 152 participants of 12,759 images with annotations at a resolution of 400×640.
- Generative data set collected with 152 participants of 252,690 images at a resolution 400×600.
- Sequence data set collected with 152 participants of 91,200 images at a resolution of 400×640, with duration of 1.5 seconds for each participant, sampled at 200 Hz.
- Left and right paired human eye corneal topography in the form of point cloud collected for 143 participants.
Announcement of the Challenge Winners
Semantic Segmentation Challenge
Accuracy (mIOU): 0.8948
Model # Parameters: 416088
First Place: Team RIT
Team Members: Aayush Chaudhary, Rakshit Kothari, Manoj Acharya, Shusil Dangi, Nitinraj Nair, Reynold Bailey, Christopher Kanan, Gabriel Diaz, and Jeff Pelz
Accuracy (mIOU): 0.9528
Model # Parameters: 248900
Second Place: Team Tetelias
Team Member: Teternikov Ilia Anatolyevich
Accuracy (mIOU): 0.9519
Model # Parameters: 242664
Third Place: Team Couger AI
Team Members: Devanathan Sabarinathan and Priya Kansal
Accuracy (mIOU): 0.949
Model # Parameters: 258021
Synthetic Eye Generation Challenge
First Place: Team AIT
Team Members: Seonwook Park, Xucong Zhang, Shalini De Mello, Otmar Hilliges, and Marcel Buehler
Second Place: Team PAU
Team Members: Yu Yu and Jean-Marc Odobez
Third Place: Team Tomcarrot
Team Member: Tom Hao Bu
How to enter the Challenge(s)
Step 1: Visit the challenge website and read the Official Rules (Rules for Semantic Segmentation Challenge; Rules for Synthetic Eye Generation Challenge) which govern your participation in each challenge.
Step 2: Submit the following information to the email address email@example.com to request access to the challenge data (“OpenEDS”):
- Job Title:
- Contact email:
- Members of Team:
By submitting your request to access OpenEDS, you agree to the Official Rules for the challenge that you are participating in. The Official Rules are a binding contract and govern your use of OpenEDS, and are linked below:
Step 3: Create an account at evalAI.cloudcv.org to use for one or both of the challenges.
Step 4: Design your model based on the training data and/or validation data available in OpenEDS.
Step 5: Generate a JSON file for results produced by your model as applied to the test dataset included in OpenEDS. The JSON file must be generated as follows:
The scripts to generate JSON files can be found in the submission_scripts folder of OpenEDS. The instruction to use the scripts and create your submission JSON are as follows:
- Save your results in an individual .npy file per test image in uint8 format
- Make a text file with each line containing the fullpath of the result .npy files. An example of the required text file is provided in the submissions_scripts folder and is named pred_npy_list.txt.
- Run the following command on terminal:
- For semantic segmentation challenge:
python create_json_ss.py --list-file <LIST FILE> --submission-json <SUBMISSION JSON> --num-model-params <YOUR MODEL PARAMS NUMBER>
- For image-synthesis challenge:
python create_json_regen.py --list-file <LIST FILE> --submission-json <SUBMISSION JSON>
- For semantic segmentation challenge:
Submissions must comply with the Official Rules of the applicable challenge.
Step 6: Login to your EvalAI account and upload your JSON file in compressed zip format to the applicable challenge portal: (1) Semantic segmentation challenge (2) Synthetic eye generation challenge. The scores will be made available on the EvalAI leadership board for each challenge respectively.
Winners of the 2019 OpenEDS Challenges will be announced on or about September 30, 2019.
- Challenge participation deadline: September 15, 2019
- Notifications to winners: September 22, 2019
- Camera ready paper deadline*: September 27, 2019
- Workshop: November 2, 2019
*Accepted challenge papers will be archived on IEEE Xplore and CVF open access.
Facebook Reality Labs
University College London
Visiting Scientist, Facebook Reality Labs
Postdoctoral Research Scientist, Facebook Reality Labs
Facebook Reality Labs
Facebook Reality Labs
Sachin S. Talathi
Facebook Reality Labs
Facebook Eye Tracking Semantic Segmentation Challenge
NO PURCHASE NECESSARY TO ENTER OR WIN A PRIZE IN THIS CONTEST. A PURCHASE WILL NOT INCREASE YOUR CHANCES OF WINNING. INTERNET ACCESS AND A VALID EMAIL ADDRESS ARE REQUIRED TO PARTICIPATE. TRAVEL TO SOUTH KOREA BETWEEN 10/27/19 AND 11/2/19 IS REQUIRED TO RECEIVE A PRIZE IN THIS CONTEST. Open only to individuals who are at least 18 and the age of majority in jurisdiction of residence and are legal residents of any area, country, state, territory, or province where applicable laws do not prohibit participating or receiving a prize in the Contest and excludes China, Kenya, Venezuela, Argentina, Denmark, Greece, Quebec, Cuba, Iran, North Korea, Sudan, Myanmar/Burma, Syria, Zimbabwe, Iraq, Lebanon, Liberia, Libya, Somalia, Zimbabwe, Belarus, Balkans, and any other area or country designated by the applicable agency that designates trade sanctions. Submission required between 12.00AM PDT on 05/03/2019 and 11:59:59 PDT on 09/15/2019. Access to data set requires request by email. Subject to OFFICIAL RULES. Limit 1 entry per person. Void where prohibited by law. Entries will be scored based on model performance and model complexity. Total ARV of all prizes in this Contest: $13,000 USD. Sponsor: Facebook Technologies, LLC, a wholly-owned subsidiary of Facebook, Inc. 1601 Willow Road, Menlo Park, CA 94025.
Facebook Eye Tracking Synthetic Eye Generation Challenge
NO PURCHASE NECESSARY TO ENTER OR WIN A PRIZE IN THIS CONTEST. A PURCHASE WILL NOT INCREASE YOUR CHANCES OF WINNING. INTERNET ACCESS AND A VALID EMAIL ADDRESS ARE REQUIRED TO PARTICIPATE. TRAVEL TO SOUTH KOREA BETWEEN 10/27/19 AND 11/2/19 IS REQUIRED TO RECEIVE A PRIZE IN THIS CONTEST. Open only to individuals who are at least 18 and the age of majority in jurisdiction of residence and are legal residents of any area, country, state, territory, or province where applicable laws do not prohibit participating or receiving a prize in the Contest and excludes China, Kenya, Venezuela, Argentina, Denmark, Greece, Quebec, Cuba, Iran, North Korea, Sudan, Myanmar/Burma, Syria, Zimbabwe, Iraq, Lebanon, Liberia, Libya, Somalia, Zimbabwe, Belarus, Balkans, and any other area or country designated by the applicable agency that designates trade sanctions. Submission required between 12.00AM PDT on 05/03/2019 and 11:59:59 PDT on 09/15/2019. Access to data set requires request by email. Subject to OFFICIAL RULES. Limit 1 entry per person. Void where prohibited by law. Entries will be scored based on model performance. Total ARV of all prizes in this Contest: $13,000 USD. Sponsor: Facebook Technologies, LLC, a wholly-owned subsidiary of Facebook, Inc. 1601 Willow Road, Menlo Park, CA 94025.