Yifan Xing

MSCV, CMU

|

GMD - GOTURN MDNet Fused Generic Object Tracker

GMD is a Siamese Convolutional Neural Network based tracker. It combines thecharacteristics of high-speed tracker GOTURN and mechanisms from MDNet for a timely feedback on model update. In particular, GMD focuses on aclassification based approach that achieves timely appearance model adaptationthrough online learning. By feeding the Siamese network the paired informationof frame t-1 target and a set of candidates sampled from frame t, the networklearns an observation model that assigns score to each of the candidates. Thescore measurement in turn provides signal for timely update of the network.Besides, ROI-pooling over the candidates is used to speed up the feed-forwardprocess. Furthermore, GMD is trained using imagenet video dataset that hashigh object appearance variations. GMD will appear in VOT2017 Challenge Workshop (held in conjunction with ICCV 2017).

Qualitative Results:
Check out the setup page on GitHub!