Abstract | ||
---|---|---|
The recent Correlation Filter (CF) based methods have shown attractive performance in visual tracking task. In real-time CF based trackers, they usually adopt hand-crafted features (e. g., HOG and Color Names), while these artificially designed features still have redundancy and can be further compressed and refined. In this paper, we design a lightweight network to offline learn how to compress the hand-crafted features for better and faster correlation tracking. To achieve this goal, we adopt CF as one layer in the network to force the learned model to be suitable for tracking task. Besides, we apply residual structure to avoid the overfitting problem in the training process. Our simple yet effective network is universal and can be applied to existing CF based trackers. After adopting our lightweight network, several state-of-the-art CF based trackers are improved in both tracking accuracy and efficiency. |
Year | DOI | Venue |
---|---|---|
2018 | 10.1007/978-3-030-00776-8_32 | ADVANCES IN MULTIMEDIA INFORMATION PROCESSING, PT I |
Keywords | Field | DocType |
Correlation tracking,Feature compression,Hand-crafted features | Compression (physics),BitTorrent tracker,Computer vision,Residual,Correlation filter,Pattern recognition,Computer science,Eye tracking,Redundancy (engineering),Correlation,Artificial intelligence,Overfitting | Conference |
Volume | ISSN | Citations |
11164 | 0302-9743 | 0 |
PageRank | References | Authors |
0.34 | 13 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Chao Xie | 1 | 10 | 4.00 |
Ning Wang | 2 | 230 | 87.46 |
Wengang Zhou | 3 | 1226 | 79.31 |
Weiping Li | 4 | 178 | 15.67 |
Houqiang Li | 5 | 2090 | 172.30 |