Deep Spatial-Temporal Field For Human Head Orientation Estimation

We present a human-head-orientation estimation approach which enables effective estimation of head orientations of multiple individuals appeared within the same scene. Our approach bases deep representation in order to obtain adequate visual features of head orientations. To boost the estimation performance, we propose a conditional random field which fuses shallow feature, deep feature and spatial-temporal contextual cues, where the fusing parameters are learned from data via structured support vector machine. We demonstrate that the three components of fusion are complementary to each other in terms of head orientation estimation. Meanwhile, the proposed spatial-temporal field outperforms the state-of-the-art significantly on public dataset.