Abstract | ||
---|---|---|
Recent years, visual object tracking has attracted more and more attention as a fundamental topic. Many deep based trackers, especially Siamese Network based trackers, have achieved state-of-the-art performance on multiple benchmarks. However, most of these trackers applied with the first frame as template throughout the tracking process. We propose a Template Attentional Siamese Network called TASNet. The core of TASNet is combining the detection results of two template frames, where the first frame extracting discriminative features and the latest frame capturing the motion changes, to enhance model tracking effect. Template-wise weights are calculated from attention mechanism to integrate the detecting results of two templates in current frame tracking. The proposed architecture is trained from end to end on the ILSVRC2015 video dataset. Our tracker operates at frame-rates real-time and achieves state-of-the-art tracking accuracy while large deformation of the object is appeared.
|
Year | DOI | Venue |
---|---|---|
2018 | 10.1145/3301506.3301544 | Proceedings of the 2018 the 2nd International Conference on Video and Image Processing |
Keywords | Field | DocType |
Siamese network, attention mechanism, discriminative features, motion change, object tracking | Computer vision,BitTorrent tracker,End-to-end principle,Computer science,Video tracking,Artificial intelligence,Template,Discriminative model | Conference |
ISBN | Citations | PageRank |
978-1-4503-6613-7 | 0 | 0.34 |
References | Authors | |
0 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Junyan Gao | 1 | 7 | 1.50 |
Zhenguo Yang | 2 | 71 | 17.57 |
Liu Wenyin | 3 | 2531 | 215.13 |