Siamese Feature Pyramid Network for Visual Tracking

Abstract

Visual tracking is an important technology of robot-assisted surgery in 5G-health. Recently, discriminative correlation filter (DCF) methods utilizing in-network feature hierarchy in convolutional neural networks (CNNs) have made state-of-art results in visual tracking. However, their models are complex, which can not run in real-time. Different from DCF methods, SiamFC (Siamese Fully Convolutional) can operate at 86 frames-per-second, while it doesn't leverage the in-network feature hierarchy. Inspired by the high speed of SiamFC and in-network feature hierarchy in CNNs, a Siamese model based on feature pyramid network is proposed to improve tracking performance. The proposed tracking algorithm can not only benefit from fine-grained spatial details in low level features, but also the semantic information in high level features. Besides, a group of ablation experiments are conducted. Without the bells and whistles, the performance improvements are visible compared to SiamFC.