Improved Generalization of Heading Direction Estimation for Aerial Filming Using Semi-supervised Regression

摘要: In the task of Autonomous aerial filming a moving actor (e.g. person or vehicle), it is crucial to have good heading direction estimation for from visual input. However, models obtained in other similar tasks, such as pedestrian collision risk analysis and human-robot interaction, are very difficult generalize task, because difference data distributions. Towards improving generalization with less amount labeled data, this paper presents semi-supervised algorithm problem. We utilize temporal continuity unsupervised signal regularize model achieve better ability. This applied both training testing phases, which increases performance by large margin. show that leveraging unlabeled sequences, required can be significantly reduced. also discuss several important details on balancing loss, making combinations. Experimental results our approach robustly outputs different types actor. The aesthetic value video improved task.