Skip to content

isohrab/Pupil-locator

Repository files navigation

Real time pupil detection in noisy images

Finding pupil location inside the eye image is the prerequisite for eye tracking. Recent state of the art method works based on algorithmic approach and tries to find pupil features and fit an ellipse. However, If the pupil be distorted by strong reflection or eyelash or eyelid (which we call it noisy image) algorithmic approaches will be failed. In this project, which was a part of my master thesis, I designed an hybrid model by inspiring YOLO, Network in Network and using Inception as the core CNN to predict the pupil location inside the image of the eye.

The images for training are noise free and the pupil is evident. The images are automatically labeled by PuRe and the ground truth is the parameter of an ellipse. I used data augmentation to increase the data quantity and more importantly simulate the noisy images such as occlusion and strong reflection.

To evaluate the model, I used publicly available datasets ExCuse, ElSe, PupilNet. The results surpass previous state of the art result (PuRe) by 9.2% and achieved 84.4%. The model's speed in an Intel CPU Core i7 7700 is 34 fps and in GPU gtx1070 is 124 fps.

Results

Dataset I II III IV V VI VII VIII IX X XI XII XIII XIV XV XVI
Accuracy(%) 90 88 88 93 97 94 81 89 91 94 97 88 83 97 77 87
Dataset XVII XVIII XIX XX XXI XXII XXIII XXIV PI PII PIII PIV PV
Accuracy(%) 96 75 46 89 89 76 100 73 83 60 69 88 82

To watch the predicted results please visit this youtube playlist

Run model

Use model to predict the pupil location in a video use this command:

python inferno.py PATH_TO_VIDEO_FILE

or you can pass 0 for camera.

Acknowledgement

This work has been accomplished as a project for my master thesis. I would like to thank Dr. Shahram Eivazi for his generous helps and Thiago Santini for providing the training data.

Citation:

We published a nice paper from my work. If you want to reference it, please use following citation:

Shaharam Eivazi, Thiago Santini, Alireza Keshavarzi, Thomas Kübler, and Andrea Mazzei. 2019.
Improving real-time CNN-based pupil detection through domain-specific data augmentation.
In Proceedings of the 11th ACM Symposium on Eye Tracking Research & Applications (ETRA ’19).
Association for Computing Machinery, New York, NY, USA, Article 40, 1–6.
DOI:https://doi.org/10.1145/3314111.3319914

About

Real time pupil detection in noisy images

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages