Gaze Point Estimation for Four-Person-Table-Meeting Using Omnidirectional Camera
Nurul Hidayah Rahim, Ryusuke Nakayama, Takeshi Kamio, Toshiharu Kosaku, Shigang Li · 2024
Group meeting around a table involves multiple people. Analyzing the communication of group meetings needs to under-stand the interaction among multiple people. In this paper, we propose a method of gaze estimation for four-person-table-meeting by using an omnidirectional camera. An omnidirectional camera is centered on a table with four sides while four participants sit at each side of the table. The advantage of this setup is that the behavior of all the participants can be observed simultaneously by using only a single camera; however, the problems caused by the presentation of an equirectangular image, distortion, and disconnectivity, must be coped with. In this paper, four perspective images are generated from an equirectangular image so that each perspective image with 90 degrees field of view covers one participant. The task of gaze estimation of group meeting is reformulated as two sub-tasks for a specific subject: a classification task to determine which generated perspective image is gazed at, and a regression task to compute gaze point position in the gazed perspective image. A neural network is developed for this goal, and the effectiveness of the proposed method is shown by the experimental results.