基于子任务学习的双分支个性化视线估计网络

Episode-based two-branch personalization network for gaze estimation

针对视线估计任务中个性化信息处理难的问题,提出了一种基于子任务学习的双分支个性化视线估计网络(EbPN)。首先,该网络包含表观信息感知分支和个性化信息感知分支,以分别刻画用户丰富的表观信息以及个性化信息;然后,构造一系列子任务以替代独立样本对模型进行训练优化,其中各子任务均模仿了一个小样本视线估计任务。该网络充分获取了用户的表观信息及个性化信息,可有效减小个性化偏差,提高视线估计模型的精度,且在没有任何校准样本的情况下,仍然具有很好的泛化性能。在GazeCapture和MPIIGaze数据集上开展的大量实验证明了EbPN的有效性。

Aiming at the challenges of personalized information processing in the task of gaze estimation,a two-branch episode-based personalization network(EbPN)is proposed.Firstly,the EbPN model,including appearance-aware network and personalization-aware network,aims to characterize abundant appearance and personalized information respectively.Secondly,the model is optimized with a collection of episodes rather than individual samples,each of which mimics a few-shot learning task.The model can take full advantage of appearance and personalized information of users,which helps eliminate personalized biases and improves the accuracy for gaze estimation model.In addition,the model has satisfactory generalization performance even without any calibration samples.Extensive experiments on GazeCapture and MPIIGaze datasets demonstrate the effectiveness of EbPN model.