PROJECT TITLE :

Deep Spatial and Temporal Network for Robust Visual Object Tracking

ABSTRACT:

For visual tracking, there are two crucial components: (a) the appearance of the object and (b) the motion of the object. Since deep learning's higher representation capacity and powerful learning ability have lately been used to augment many current techniques, most of which use object appearances but few of which utilise object motions, many existing methods have recently turned to deep learning. We have developed a deep spatial and temporal network (DSTN) for visual tracking that takes advantage of the object representations from each frame and their dynamics over numerous frames in a video to provide compact object appearances and record temporal variations effectively.. DSTN can detect the minor differences in the target's spatial and temporal variations and hence benefits from both off-line training and on-line fine-tuning when implemented in a tracking pipeline in a coarse-to-fine manner. Additionally, we have tested our DSTN method on the OTB-2013 and OTB-2015 benchmarks as well as the VOT2015 and VOT2017 benchmarks and our results show that our method can compete with the state-of-the-art methodologies. In order to aid future research into this issue, all of the work's source code, trained models, and experimental findings will be made publicly available.