Fig. 3: Architecture, performance, and attention distribution of our deep learning algorithm.
From: Human and algorithmic visual attention in driving tasks
a Architecture of our deep learning algorithm. b Model accuracy of the non-pretrained algorithm with different model sizes. c Model accuracy with and without pretraining. Dashed line indicates human averages in each task. d Examples of the algorithm’s attention distribution as finetuning progressed. Red squares indicate areas of interest (The real-life driving scenes are selected from the DRAMA and BDD-OIA datasets.).
