DasiamRPN论文阅读
文献: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]
文章主要贡献
1.训练数据的扩充
- 加入Detection pair (ImageNet,COCO中做数据增广)
- 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,将Proposal与exemplar的相似性度量得分减去所有之前预先得到的Distrator(NMS将网络提出的proposal去冗余,去掉classification score最高的proposal,在剩下的Distrator set 中保留score大于给定阈值的proposal)与当前proposal 的score(相似性度量)的加权和的平均
3.long term Tracking
- 当追丢时用local to global stategy 恒定step迭代的增加搜索区域的大小
具体而言
1. 扩充数据集
![]() | ![]() | ![]() |
---|---|---|
detection pairs | negative pair from same categoriess | negative pairs from different categories |
2. Distractor Model
TEST:
Γ
(
n
)
=
(
n
−
1
)
!
∀
n
∈
N
\Gamma(n)=(n-1)!\quad\forall n\in\mathbb N
Γ(n)=(n−1)!∀n∈N
传统的SiamTracking是用求相似性度量用以下公式:
f
(
x
)
=
φ
(
x
)
∗
φ
(
z
)
+
b
⋅
1
f(x)=\varphi(x)*\varphi(z)+b\cdot\mathbf{1}
f(x)=φ(x)∗φ(z)+b⋅1
- 作者提出将NMS将网络提出的proposal去冗余,去掉classification score最高的proposal,在剩下的Distrator set 中保留score大于给定阈值的proposal)与当前proposal 的score(相似性度量)的加权和的平均
q = arg max p k ∈ P f ( z , p k ) − α ^ − ∑ i = 1 n α i f ( d i , p k ) ∑ i = 1 n α i 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}} q=pk∈Pargmaxf(z,pk)−∑i=1nαiα^−∑i=1nαif(di,pk)
P \mathcal{P} P 是score在top-k的proposal, α i \alpha_{i} αi是每个干扰proposal的权重(paper中是全为1), d i d_{i} di是第 i i i 个 distractor proposal - 因为自相关操作是线性的,则将 φ ( p k ) \varphi(p_{k}) φ(pk)提出来:
q = arg max p k ∈ P ( φ ( z ) − α ^ ∑ i = 1 n α i f ( d i , p k ) ∑ i = 1 n α i ) ∗ φ ( p k ) 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}) q=pk∈Pargmax(φ(z)−∑i=1nαiα^∑i=1nαif(di,pk))∗φ(pk)
3.Long term Tracking
- 当追丢时用local to global stategy 恒定step迭代的增加搜索区域的大小