In recent years, pupil detection in human eye images or videos has played a key role in many fields. In the field of eye tracking, the position of the center of the pupil is a basic problem, and the error of pupil detection will be magnified in subsequent calculations, which will seriously affect the performance of eye tracking. In this paper, we propose to use the currently popular semantic segmentation network for pupil detection task. We first train the Unet architecture as a benchmark, then introduce two different attention modules into Unet, and compare with the benchmark network. The results show that our method has a higher detection rate within 1-15 pixel errors. We also added an ellipse fitting error term to the loss function of the network to further improve the network performance. The training of the model is done on the LPW dataset. Finally, we also investigate the effect of data augmentation on generalization performance, with the model trained on the LPW dataset and tested on the I-SOCIALDB dataset. Although data enhancement slightly reduces the detection rate of the model in the original data set, it can improve the generalization performance of the model. © COPYRIGHT SPIE. Downloading of the abstract is permitted for personal use only.