RITnet：用于注视跟踪的实时眼部语义分割网络
计算机视觉与模式识别
2020-12-22 v1 图像与视频处理
摘要
准确的眼部分割可改善视线估计并支持基于视觉注意的交互式计算；然而，现有的眼部分割方法存在诸如因人而异的精度、缺乏鲁棒性以及无法实时运行等问题。在此，我们提出 RITnet 模型，这是一种结合 U-Net 与 DenseNet 的深度神经网络。RITnet 大小低于 1 MB，在 2019 OpenEDS 语义分割挑战赛上达到 95.3% 的精度。使用 GeForce GTX 1080 Ti，RITnet 以 >300Hz 跟踪，可实现实时注视跟踪应用。预训练模型与源代码见 https://bitbucket.org/eye-ush/ritnet/。
引用
@article{arxiv.1910.00694,
title = {RITnet: Real-time Semantic Segmentation of the Eye for Gaze Tracking},
author = {Aayush K. Chaudhary and Rakshit Kothari and Manoj Acharya and Shusil Dangi and Nitinraj Nair and Reynold Bailey and Christopher Kanan and Gabriel Diaz and Jeff B. Pelz},
journal= {arXiv preprint arXiv:1910.00694},
year = {2020}
}
备注
This model is the winning submission for OpenEDS Semantic Segmentation Challenge for Eye images https://research.fb.com/programs/openeds-challenge/. To appear in ICCVW 2019. ("Pre-trained models and source code are available https://bitbucket.org/eye-ush/ritnet/.")