Re3 : Real-Time Recurrent Regression Networks for Visual Tracking of Generic Objects

Re3 : Real-Time Recurrent Regression Networks for Visual Tracking of Generic Objects

2019-10-04 14:42:54

Paper：https://arxiv.org/pdf/1705.06368.pdf

Code（TensorFlow）：https://gitlab.com/danielgordon10/re3-tensorflow

Demo Video: https://www.youtube.com/watch?v=RByCiOLlxug&feature=youtu.be

Related Trackers: GOTURN [Blog]

This paper is developed based on deep regression network, the key idea is to utilize the LSTM network to memorize the history information (i.e. the tracking results). In addition, they also utilize multi-layer's features for better target representation. The overall pipeline can be found in following figure:

This tracker work well, but the training time need to take about one week (7 days), as it mentioned in this paper. It is because the original deep regression network is already hard to train, and the introduced LSTM further increased the difficulty. The efficiency is really fast, 150 FPS on a GPU.

The imporvement based on GOTURN is significant, according to their experiments, as shown in Fig. 4.

Although this tracker seems simple, but it really works well in some videos and fast enough for practical applications, such as uav, robotic.

巴特西

Re3 : Real-Time Recurrent Regression Networks for Visual Tracking of Generic Objects

最新文章

热门文章