During video-conferencing, there is an inherent inability for participants to be engaged eye-to-eye due to disparities between the physical locations of the webcam and the screen. Participants would normally appear as if they are looking down on each other instead of directly at each other. Real-time gaze correction systems have therefore emerged to give video conferencing participants the illusion of eye contact.
Our work attempts to improve a previous gaze correction system which adopts eye replacement strategy. Under the mentioned system, the participants would feel like the eye gaze of the other end user has always been unnaturally locked. We take a closer look into when and how the system should perform eye replacement, which could respect the actual intentions of users and suppress visual artifacts during transition (replacement and non-replacement). We conduct small-range user studies to analyze their feedbacks and enable the system to understand user intentions based on eye appearances. We also implement the smoothing component to mitigate abrupt changes and unnatural artifacts during the transition period. Our work is an effort to make the current gaze correction system more adaptable to complicated scenarios in real video conferencing.