.png){kind=icon}
.png){kind=icon}
Return
Transfer the global knowledge for current gaze estimation
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-023-17484-2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The Gaze Estimation (GE) task aims to estimate the gaze direction of the current frame with the existing image information. We find a reasonable scene where we only can get the information of history frames for testing while we can get the information of the history and future frames for training. This is because that testing is at a current moment, while training is based on data collected in the past. To adapt to this scene, the previous methods only use the current frame or the history frames to estimate the gaze of the current frame. Unlike them, we think that the future frames of the training phase can improve the model's performance in the testing phase. So, we propose a novel framework of teacher-student learning to transfer the global knowledge of history and future frames from the teacher network to the student network and call it TGKF (Transfer Global Knowledge Framework). Specifically, we use the history and future frames as the input of the teacher network and mine the global knowledge from historical and future perspectives. We use the history frames as the input of the student network and mine the knowledge from the historical perspective. Then, we transfer the global knowledge from the teacher network to the student network by knowledge distillation. In this way, the student network has the ability to infer global knowledge from historical knowledge. Extensive experiments and ablation studies have shown the effectiveness of TGKF. With our proposed TGKF, we achieve favorable results on two benchmark datasets.
Keywords:
Gaze estimation
Teacher-student learning
Knowledge distillation
Temporal relation modeling
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W