SiamRPN.md

High Performance Visual Tracking with Siamese Region Proposal Network

cvpr zhihu pytorch SiamRPN算法汇总

Introduction

1. tracker分类
   1. 基于correlation filter
      1. 同时online跟踪并更新filter的参数
      2. 可以基于传统feature或deep feature
      3. model更新影响跟踪速度
   2. 基于很强的deep feature
      1. 不需要 model更新
      2. 精度受到data domain影响
2. This paper：把跟踪看做one-shot detection

Framework

1. 一个空间位置上有$k$个anchor
2. 分类输出size为$wh2k$
3. 回归输出size为$wh4k$

Tracking as one-shot detection

one-shot detection：仅仅根据初始信息来定位 Because the local detection task is based on the category information only given by the template on initial frame, it can be viewed as one-shot detection

Anchor选择

1. 排除太远的框
2. 用 cosine window 和从尺度变化来re-rank the proposals’ score