Easy One-Stream Transformer Tracker

To replace the traditional two-stream tracking paradigm, exploiting a one-stream tracking architecture has recently drawn extensive interest. But redundant computations occur as not all tokens are attentive in MHSA, introducing the noise brought by inattention information interaction and matching. To address the aforementioned issue, we introduce an easy one-stream transformer track(EOTrack), which reorganizes search tokens by retaining attentive ones and merging inattentive ones. This approach accelerates subsequent computations, enhancing the tracker's discrimination against the target by gradually mitigating noise. Additionally, we eliminate background information when updating templates without increasing training costs to provide diversified and high-quality positive samples for the tracker. Our designed easy one-stream framework is effective and concise, and its effectiveness has been proven on three datasets.

I. Introduction

Object tracking in computer vision poses a fundamental and formidable challenge, involving the continuous localization of a target's state across successive frames, based on the initial frame's annotation. This task holds significant potential with applications spanning visual surveillance and augmented reality.