基于双帧图模型的视频物体分割

通过将空间及时序信息有效结合, 提出基于双帧图模型的单目视频物体分割算法。首先, 通过手工交互得到初始帧的分割, 并据此训练获取前景物体及背景的颜色模型。然后, 利用双帧图模型融合当前帧的颜色信息以及由颜色差得到的空间和时序约束。根据观测到的颜色差异与物体运动的线性关系, 提出运动自适应的时序约束因子, 它能够随视频中物体运动变化自适应地调节。最后, 通过二值图割法计算当前帧的分割结果并更新颜色模型。利用双帧图模型可循序地对视频中的下一帧进行分割。实验结果证实, 提出的自适应时序约束因子可以提高物体分割结果的准确性和时序一致性, 量化指标表明此算法在视频物体分割中可获得更优结果。

Video Object Segmentation via Two-Frame Graph Model

The authors present a novel algorithm for object segmentation in monocular video sequences, which is based on proposed two-frame graph model. First, the bi-layer color models can be learnt from the interactive segmentation on the first frame. To sequentially extract the object on the next frame, the two-frame graph model is proposed, in which the color cues and spatio-temporal constraints are combined. Based on the observation that the amount of the object movement is approximate linearly related to the summation of the temporal contrasts between adjacent frames, the motion-adaptive temporal-constraint factor is proposed in order to control the weight of temporal constraints in the graph model. The simplification of the graph model increases the efficiency of the algorithm. Finally, each frame is segmented by binary graph cut and the color models are updated. Experimental results show that the adaptive temporal-constraint factor can improve the accuracy and temporal coherence of the results. Experiments on the public database indicate that the proposed algorithm is more efficient and effective.