文獻:DaSiamRPN: Zheng Zhu, Qiang Wang, Bo Li, Wu Wei, Junjie Yan, Weiming Hu."Distractor-aware Siamese Networks for Visual Object Tracking." ECCV (2018). [paper][github]
文章主要貢獻
1.訓練數據的擴充
- 加入Detection pair (ImageNet,COCO中做數據增廣)
- negative simple in same categories (Called Distractor-aware Training)
- negative simple in different categories (Called Distractor-aware Training)
2.Distractor Model
- 引入Distractor Model,將Proposal與exemplar的相似性度量得分減去所有之前預先得到的Distrator(NMS將網絡提出的proposal去冗余,去掉classification score最高的proposal,在剩下的Distrator set 中保留score大於給定閾值的proposal)與當前proposal 的score(相似性度量)的加權和的平均
3.long term Tracking
- 當追丟時用local to global stategy 恆定step迭代的增加搜索區域的大小
具體而言
1. 擴充數據集
![]() |
![]() |
![]() |
---|---|---|
detection pairs | negative pair from same categoriess | negative pairs from different categories |
2. Distractor Model
TEST:\(\Gamma(n)=(n-1)!\quad\forall n\in\mathbb N\)
傳統的SiamTracking是用求相似性度量用以下公式:
\[f(x)=\varphi(x)*\varphi(z)+b\cdot\mathbf{1} \]
- 作者提出將NMS將網絡提出的proposal去冗余,去掉classification score最高的proposal,在剩下的Distrator set 中保留score大於給定閾值的proposal)與當前proposal 的score(相似性度量)的加權和的平均
\[q=\mathop{\arg\max}\limits_{p_{k}\in\mathcal{P}} f(z,p_{k})- \frac{\hat{\alpha}-\sum_{i=1}^{n}\alpha_{i}f(d_{i},p_{k})} {\sum_{i=1}^{n}\alpha_{i}}\]
\(\mathcal{P}\) 是score在top-k的proposal, \(\alpha_{i}\)是每個干擾proposal的權重(paper中是全為1), \(d_{i}\)是第 \(i\) 個 distractor proposal
- 因為自相關操作是線性的,則將\(\varphi(p_{k})\)提出來:
\[q=\mathop{\arg\max}\limits_{p_{k}\in\mathcal{P}}(\varphi(z)-\frac{\hat{\alpha}\sum_{i=1}^{n}\alpha_{i}f(d_{i},p_{k})} {\sum_{i=1}^{n}\alpha_{i}})*\varphi(p_{k})\]
3.Long term Tracking
- 當追丟時用local to global stategy 恆定step迭代的增加搜索區域的大小