.png){kind=icon}
.png){kind=icon}
Return
Learning Motion-Perceive Siamese network for robust visual object tracking
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patrec.2023.07.011.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Siamese networks enable end-to-end training for visual tracking, achieving excellent performance in recent years. However, classical Siamese-based formulation relies on an offline-trained appearance model to perform tracking for each frame, ignoring the temporal variation at the online stage. As been verified that temporal motion is crucial for robust and accurate tracking, we propose a novel Motion-Perceive Siamese network (SiamMP) that explicitly predicts motion patterns, providing complementary clues for the appearance-only formulation. Specifically, successive historical frames are collected, with their appearance and potential trajectory being employed to predict the next state, achieving motion awareness correspondingly. Besides, an adaptive fusion module is dedicated to performing a decision-level negotiation between the tracking evidence of the appearance and the motion models. To verify the effectiveness and merit of our SiamMP, extensive experiments are conducted on several challenging benchmarks, including LaSOT, OTB100, GOT-10k, TC128, and DTB70. The comparison and analysis of the obtained results demonstrate the necessity and validity of involving the motion-perceive design in the Siamese framework.
Keywords:
Visual object tracking
Siamese network
Motion prediction
Adaptive fusion