Generic Pixel Level Object Tracker Using Bi-Channel Fully Convolutional Network

As most of the object tracking algorithms predict bounding boxes to cover the target, pixel-level tracking methods provide a better description of the target. However, it remains challenging for a tracker to precisely identify detailed foreground areas of the target. In this work, we propose a novel bi-channel fully convolutional neural network to tackle the generic pixel-level object tracking problem. By capturing and fusing both low-level and high-level temporal information, our network is able to produce pixel-level foreground mask of the target accurately. In particular, our model neither updates parameters to fit the tracked target nor requires prior knowledge about the category of the target. Experimental results show that the proposed network achieves compelling performance on challenging videos in comparison with competitive tracking algorithms.