SIGHT LINE TRACKING MODEL TRAINING METHOD, AND SIGHT LINE TRACKING METHOD AND DEVICE

摘要

Disclosed in the present application is a sight line tracking model training method, comprising: obtaining a training sample set; processing an eye sample image by means of an initial sight line tracking model to obtain a predicted sight line vector of the eye sample image; determining model loss according to a cosine distance between the predicted sight line vector and a marked sight line vector; and performing iteration adjustment on reference parameters of the initial sight line tracking model until the model loss satisfies a convergence condition to obtain a target sight line tracking model. By adopting the solution provided in the present application, performing sight line tracking does not need to use peripherals, and only needs to input the collected eye image into target sight line tracking, so that a procedure of sight line tracking is simplified, and by using the cosine distance as model loss to train the model, a difference between the predicted value and the marked value can be better represented, thereby improving the prediction accuracy of the sight line tracking model obtained by training.