Abstract:Recent breakthroughs of neural network structures,such as stacked hourglass network and heat-map regression,led the accuracy of face landmark localization to saturate to an extraordinary level.However,a new problem of landmark jittering emerges when applying these methods to video frames directly.Landmark jitter is a phenomenon of undesired fitful movements of certain landmark point while the actual motion of the face,relative to the frames,is along a regular continuous curve.Eliminating landmark jitter and providing smooth landmark sequences becomes a new challenge for video-based face alignment applications.In this paper,we propose a novel technique called ″adversarial training with SMG(Soft Mesh Grid) reverse transformation″.The network model trained with this modification can be applied straight to video streams and it will output stable predictions without deteriorating accuracy.The smoothness of our model's predictions surpasses state of the art.
何卓骋,李京. 对抗训练在人脸关键点序列稳定化问题中的应用[J]. 小型微型计算机系统, 2021, 42(7): 1407-1414.
HE Zhuo-cheng,LI Jing. Adversarial Training Based Stabilizing Mechanism for Real-time Face Landmark Localization. Journal of Chinese Computer Systems, 2021, 42(7): 1407-1414.